Responses of neurons in the primate taste cortex to glutamate
Autonomic control of the eye眼的自主神经
Autonomic Control of the EyeDavid H.McDougal1and Paul D.Gamlin*2ABSTRACTThe autonomic nervous system influences numerous ocular functions.It does this by way of parasympathetic innervation from postganglionicfibers that originate from neurons in the ciliary and pterygopalatine ganglia,and by way of sympathetic innervation from postganglionicfibers that originate from neurons in the superior cervical ganglion.Ciliary ganglion neurons project to the ciliary body and the sphincter pupillae muscle of the iris to control ocular accommodation and pupil constriction,respectively.Superior cervical ganglion neurons project to the dilator pupillae muscle of the iris to control pupil dilation.Ocular bloodflow is controlled both via direct autonomic influences on the vasculature of the optic nerve,choroid,ciliary body,and iris,as well as via indirect influences on retinal bloodflow.In mammals,this vasculature is innervated by vasodilatoryfibers from the pterygopalatine ganglion,and by vasoconstrictivefibers from the superior cervical ganglion.Intraocular pressure is regulated primarily through the balance of aqueous humor formation and outflow.Autonomic regulation of ciliary body blood vessels and the ciliary epithelium is an important determinant of aqueous humor formation;autonomic regulation of the trabecular meshwork and episcleral blood vessels is an important determinant of aqueous humor outflow.These tissues are all innervated byfibers from the pterygopalatine and superior cervical ganglia.In addition to these classical autonomic pathways,trigeminal sensory fibers exert local,intrinsic influences on many of these regions of the eye,as well as on some neurons within the ciliary and pterygopalatine ganglia.C 2015American Physiological Society. Compr Physiol5:439-473,2015.IntroductionThe ocular projections of the autonomic nervous system influ-ence numerous functions of the eye.These include:(i)pupil diameter and ocular accommodation(ACC),which are con-trolled by the intrinsic muscles of the eye located in the iris and ciliary body,respectively—these structures are innervated by postganglionicfibers from the ciliary(parasympathetic) and superior cervical(sympathetic)ganglia;and(ii)ocular bloodflow,which is controlled via innervation of the vascu-lature within the optic nerve,the retina,choroid,ciliary body, and iris.In mammals,these vascular beds are innervated by postganglionicfibers from the pterygopalatine(parasympa-thetic)and superior cervical(sympathetic)ganglia.In birds, the ciliary ganglion also contributes to the parasympathetic innervation of the choroid,and there may be a small con-tribution from this ganglion in mammals;(iii)intraocular pressure(IOP),which is regulated primarily through changes in aqueous humor formation and outflow.Autonomic regu-lation of ciliary body blood vessels and the ciliary epithe-lium are important determinants of aqueous humor forma-tion,while regulation of the trabecular meshwork and epis-cleral blood vessels are important determinants of aqueous humor outflow—these structures are innervated by postgan-glionicfibers from the pterygopalatine(parasympathetic)and superior cervical(sympathetic)ganglia(Fig.1).However,as discussed below,these functional subdivisions are somewhat arbitrary since they are often interrelated and do not operate in isolation.For example,both iris and ciliary muscle contraction can influence aqueous humor outflow,while alterations in choroidal bloodflow will cause changes in IOP.The information presented below focuses not only on the autonomic innervation of the eye,but also the central auto-nomic pathways controlling pupillary reflexes,ACC,ocular bloodflow,and IOP.Our discussion of these topics is concen-trated on primates,including humans,but includes data from other species where it suggests alternative connections that have yet to be investigated in primates,or where primate data is unavailable.Unless otherwise stated,primary references reflect studies conducted in the primate model.A comprehen-sive list of abbreviations used in this article can be found in Table1.As shown in Figure1,parasympathetic innervation of the eye arises from two sources:(i)neurons in the Edinger-Westphal preganglionic(EWpg)cell group,the autonomic subdivision of the third cranial nerve nucleus,which lies in the rostral mesencephalon;(ii)preganglionic neurons in the superior salivatory nucleus(SSN),the parasympathetic com-ponent of the seventh cranial nerve nucleus,which lie in the *Correspondence to pgamlin@1Neurobiology of Metabolic Dysfunction Laboratory,Pennington Biomedical Research Center,USA2Department of Ophthalmology,University of Alabama at Birmingham,USAPublished online,January2015() DOI:10.1002/cphy.c140014Copyright C American Physiological Society.Autonomic Control of the Eye Comprehensive PhysiologyEWpgSSN PPGIML SCGSphincter pupillae Ciliary muscleDilator pupillae Sphincter pupillae Ciliary muscleDilator pupillae Choroidal blood vessels*Trabecular meshwork Choroidal blood vessels Ciliary epithelium Ciliary body blood vessels Iris blood vesselsACh VIPNA NPY ATPOptic nerve head Choroidal blood vessels Iris blood vesselsOptic nerve head Episcleral blood vesselsTrabecular meshwork Ciliary epithelium Ciliary body blood vessels Episcleral blood vesselsOphthalmic, central retinal, ciliary arteries Ophthalmic, central retinal, ciliary arteries Figure 1A schematic diagram showing the parasympathetic andsympathetic innervation of the eye.Neurotransmitters and neuropep-tides that are generally present in postganglionic neurons are iden-tified.∗Only seen in avian studies to date.Abbreviations:ACh,Acetylcholine;ATP,Adenosine triphosphate;CG,Ciliary ganglion;EWpg,Edinger-Westphal nucleus,preganglionic;IML,Intermediolat-eral nucleus (cell column);NA,Noradrenaline;nNOS,neuronal nitric oxide synthase;NPY,Neuropeptide Y;PPG,Pterygopalatine ganglion;SCG,Superior cervical ganglion;SSN,Superior salivatory nucleus;VIP,Vasoactive intestinal peptide.medulla oblongata.The neurons in EWpg project by way of the oculomotor (III)nerve to postganglionic cells in the ciliary ganglion.Neurons in the SSN project by way of the greater petrosal branch of the facial (VII)nerve to postgan-glionic cells in the pterygopalatine ganglion (also termed the sphenopalatine ganglion).Sympathetic innervation of the eye arises from preganglionic neurons located in the C8-T2seg-ments of the spinal cord,a region termed the ciliospinal center of Budge (and Waller).The axons of these preganglionic neu-rons project to the sympathetic chain ganglia and travel in the sympathetic trunk to the superior cervical ganglion where they contact postganglionic neurons (148,178).The axons of these postganglionic neurons project from the superior cer-vical ganglion to the orbit,where they enter the eye via the short and long ciliary nerves,as well as through the optic canal (322).In addition to these classical,extrinsic autonomic con-trol pathways,it is clear that there are local,intrinsic influ-ences exerted by trigeminal sensory fibers on many regions of the eye,as well as on some neurons within the ciliary and pterygopalatine ganglia (Fig.2).Both sympathetic and parasympathetic fibers are also present in the mammalian cornea.However,their density varies significantly between species (254).In rabbits and cats,sympathetic innervation is quite dense (80,189,203,230,251,386).However,in pri-mates,including humans,sympathetic innervation appears to be either absent or very sparse (79,81,230,371,393).In addi-tion,although some parasympathetic innervation of the corneaTable 1List of AbbreviationsACC Accommodation ACCV Accommodative velocity AP Area pretectalis ATP Adenosine triphosphate cFN Caudal fastigial nucleus CGRP Calcitonin gene-related peptide CCK Cholecystokinin DAG DiacylglycerolEVP Episcleral venous pressure EW Edinger-Westphal nucleusEWcp Edinger-Westphal centrally projecting cells EWm Medial (choroidal)subdivision of the avian Edinger-Westphal nucleusEWpg Edinger-Westphal preganglionic cells HRP Horseradish peroxidase IML Intermediolateral cell IOP Intraocular pressure IP 3Inositol triphosphateipRGCs Intrinsically photosensitive retinal ganglion cellsL-NAMENitro-L-arginine methylesterm1,m2,m3...m1,m2,m3...muscarinic receptor subtypes NADPHd Nicotinamide adenine dinucleotide phosphate-diaphorase nNOS Neuronal nitric oxide synthase NPY Neuropeptide YOpn4−/−Melanopsin-deficient knockout mice PIN Posterior interposed nucleus PIPR Postillumination pupil response PLC Phospholipase C PLR Pupillary light reflex PNR Pupillary near response PON Pretectal olivary nucleus rd/rd cl Rodless/coneless knockout mice RPE Retinal pigment epithelium SCN Suprachiasmatic nucleus SOA Supraoculomotor area SP Substance PSSN Superior salivatory nucleus VIP Vasoactive intestinal peptide WGAWheat germ agglutininComprehensive Physiology Autonomic Control of the EyeCiliary bodyI r i sC o r n e aRetina RPERetina RPENVSMICNBVChoroidTGChoroidScleraP P GS C GScleraCiliaryganglion (CG)T rigeminal ganglion (TG)EWpgPonsSp5SSNIMLPterygopalantine ganglion (PPG)Superior cervical ganglion (SCG)L e n sFigure 2Autonomic and trigeminal ocular projections.The dotted line shows a ciliary ganglion projection to the choroid which,to date,has been shown only in birds.Line thickness indicates relative strength of projection.Abbreviations:BV,Blood vessels;CNS,Central nervous system;EWpg,Nucleus of Edinger-Westphal,preganglionic division;ICN,Intrinsic choroidal neurons;IML,Intermediolateral nucleus;NVSM,Nonvascular smooth muscle;SSN,Superior salivatory nucleus.has been reported in rat and cat,the existence of such innerva-tion has not been studied in other mammalian species (254).Therefore,the autonomic innervation of the cornea will not be discussed further.A detailed description of the autonomic nervous system in general is provided by Hockman (148)and Robertson (311).Also,the autonomic control of the eye and iris in nonmammalian species has recently been reviewed (260),as has the control of ocular blood flow (182,305).Autonomic Innervation of the EyeThe third nerve pathways innervating the eye The Edinger-Westphal nucleus—preganglionic neuronsThe parasympathetic,third nerve pathway originates from preganglionic neurons in the Edinger-Westphal nucleus (EW)and projects via the third cranial nerve to the ciliary ganglion.In most primates,the EW is a distinct nucleus lying imme-diately dorsal to the somatic subdivisions of the oculomotor complex.It was first described in a developmental study of human neuroanatomical material by Edinger (77)and,a short time afterward,in a neuropathological study by Westphal (418).While both authors recognized this nucleus as being cytoarchitecturally distinct from the oculomotor nucleus,the study by Westphal led to the suggestion that the EW was involved in the innervation of the iris and possibly other intraocular muscles.The primate EW is composed of relatively large spher-ical and ovoid cells (approximately 25-40μm in diameter)and other spindle-shaped neurons (15-30μm in diameter)(e.g.194,412).The preganglionic neurons within the EW were initially identified by a retrograde degeneration study (412).Subsequently,these neurons were identified by ret-rograde neuroanatomical tracer studies following injections into the ciliary ganglion of either horseradish peroxidase (HRP)(47,60),[125I]wheat germ agglutinin (WGA)(4),Autonomic Control of the Eye Comprehensive PhysiologyWGA-HRP(372),orfluorescent tracers(161).All of these studies reported that labeled,preganglionic neurons are gen-erally the larger,more spherical neurons of the EW and are only slightly smaller than the somatic motoneurons of the oculomotor complex.Based on the number of retrogradely labeled cells and the estimated number of parasympathetic axons in the oculomo-tor nerve,Burde and Loewy(47)suggested that there were 300to400preganglionic neurons in EW.However,based on the results of Akert and colleagues(4),Ishikawa and col-leagues(161),and upon personal observations,the number of preganglionic neurons in the primate EW is more likely to be in the range of800to1200.Additional studies have confirmed in many vertebrate classes that the EW is the preganglionic,parasympathetic component of the oculomotor nuclear complex,and is the central source of parasympathetic innervation of the iris,the ciliary body,and certain additional intraocular muscles and tissues(see194for a review).However,in many species including humans,some or all cells of the cytoarchitecturally defined EW are not preganglionic neurons,but are instead centrally projecting peptidergic(often urocortin-1)neurons, while the preganglionic neurons are actually located outside of the cytoarchitecturally defined EW(e.g.103,151,194,238). For example,in one of the clearest cases of this mismatch, cells in the classically defined EW of cats project to the spinal cord and cerebellum(48,216,217,313,369),while the ciliary ganglion receives its central input from a collection of pregan-glionic cells along the midline of the rostral mesencephalon and in the ventral tegmental area(197,370,398).As discussed in a recent review(194),this has led to much confusion in the literature,and it is now generally accepted that the term EW must be supplemented with a terminology based on neuronal connectivity.Specifically,the cholinergic,preganglionic neu-rons supplying the ciliary ganglion are now termed the EWpg population,while the centrally projecting,peptidergic neu-rons are termed the Edinger-Westphal centrally projecting (EWcp)population(194).This new nomenclature and the relative location of the EWpg and EWcp cell populations for various species,including humans,are shown in Figures3 and4.The ciliary ganglionThe ciliary ganglion in humans is approximately3mm in size,and located2to3mm posterior to the globe and lateral to the optic nerve.Within the ciliary ganglion,post-ganglionic neurons are contacted by the cholinergic,nico-tinic synapses of parasympathetic preganglionic axon ter-minals that in macaque,generally contain neuropeptide Y (NPY)(Grimes et al.,1998).The axons of these postgan-glionic neurons leave the ciliary ganglion to enter the eye via the short ciliary nerves.Once in the eye,postganglionic axons project in the suprachoroid space to extensively inner-vate the sphincter pupillae and ciliary muscles and,in birds, the choroidal vasculature(103).They also make a limited(A)(B)(C)(D)(E)(F)(G)(H)(I)(K)(L)(M)(N)(O)(P)Figure3Line drawings showing the organization of EWpg and EWcp in several selected species:avian,rat,cat,monkey,and human. Representative rostral(left column),middle(middle column),and cau-dal(right column)sections are shown.The EWcp is indicated by light gray shading and EWpg is indicated by dark gray shading.Scattered cells located outside the nuclear boundaries are indicated by appro-priately shaded circles.(Adapted,with permission,from Kozicz et al., 2011)(194).number of synapses on sympathetic terminals in the dila-tor pupillae muscle(173).Although postganglionic neu-rons in the ciliary ganglion receive input primarily from the preganglionic neurons of EW,there is evidence for addi-tional neuronal inputs that may act to modulate this signal (240).For example,there is good evidence that some(∼10-20%)postganglionic neurons receive calcitonin gene-related peptide/substance P(CGRP/SP)-positive trigeminal inputsComprehensive Physiology Autonomic Control of the EyeEWpgEWpgEWpgEWpgEWpgNcl. Perlia RostralHuman CaudalIII(A)(B)(C)(D)(E)IIIIIIIIIIIIIII IIIIIIEWcpEWcpEWcpEWcpEWcp EWcpCCCCCC EWcpAvianRatCatMonkeyFigure4Three dimensional(3-D)representations of human, macaque,cat,rat and pigeon oculomotor complex,to illustrate the3-D organization of EWpg and EWcp.The3-D models are cut at selected points to illustrate how frontal sections through this level would look. In cases where the population is scattered,and so not contained in a discrete nucleus,dots are used.(Adapted,with permission,from Kozicz et al.,2011.)(194).(Fig.2)(e.g.,186).Furthermore,ultrastructural analysis of the macaque ciliary ganglion via electron microscopy demon-strated heterogeneity in cellular and synaptic structure,as well as postsynaptic vesicular content(240).Therefore,the ciliary ganglion should not be considered only as a simple relay of preganglionic inputs from EW to the eye,but also as a site of potential neural integration(103).In addition,although many studies have reported that all preganglionic neurons synapse in this ganglion in mammals(e.g.,197,323),there has been some debate as to the existence of a synapse in this ganglion for both pupil-related and ACC-related pregan-glionic neurons.For example,based on physiological studies (417)and on retrograde anatomical studies(163,284),it has been reported that EW neurons do not synapse in the ciliary ganglion,but instead have a direct projection to the ciliary muscle.Other reports have disagreed with these studies and there is a growing consensus for the existence of a synapse between the pre-and postganglionic neurons in the primate ciliary ganglion(see320for a review).It appears likely that some retrograde tracer experiments showed apparent direct connections between the preganglionic neurons of the EW and their eventual peripheral targets because the tracer was taken up by preganglionicfibers that contact intraocular gan-glion cells contained within the accessory ciliary ganglia of the primate and other mammals.In some mammals these accessory ganglia are located immediately behind the sclera, while in others they are located in the suprachoroid lamina along the intraocular portions of the ciliary nerves from the scleral canal to the iris and ciliary body(197,198).Injections in the vicinity of these intraocular ganglion cells would have involved preganglionicfibers from EW neurons and would thus have resulted in retrograde labeling of EW neurons. Seventh nerve pathways innervating the eye Superior salivatory nucleus—preganglionic neurons The parasympathetic,seventh nerve pathway innervating the eye originates from preganglionic neurons in the SSN,which are located in the ventrolateral medulla,slightly dorsolateral to the facial motor nucleus.In rats,the preganglionic neurons are somewhat intermingled with,and surrounded by neurons of the A5catecholamine cell group(Fig.5)(e.g.,69,209). All SSN neurons are cholinergic,and in rabbits and rats many have also been shown to contain neuronal nitric oxide syn-thase(nNOS)(69,438,439).In rats,it has been established that one population of preganglionic SSN neurons projects by way of the greater petrosal branch of the facial nerve to the pterygopalatine ganglion(65,69,261,263,305,350,396), which sends postganglionicfibers to ocular structures which are involved in the regulation of bloodflow and IOP(see below).The other population of preganglionic neurons within the SSN are not involved in the autonomic control of the eye.These neurons project by way of the chorda tympani nerve to the submandibular ganglion(65,166,261,263,305) which sends postganglionicfibers to the submandibular andAutonomic Control of the Eye Comprehensive PhysiologySSNmlf–11 mmtsGiARPapyRMgLPGiPPym7g7pr VeMVeLdas DCicpscpVeCbAcs7rs sp5sp58nFigure 5Diagram showing the location of prechoroidal SSN neu-rons in rat.Abbreviations:8n,vestibulocochlear nerve;Acs7,acces-sory facial nucleus;das,dorsal acoustic stria;DC,dorsal cochlear nucleus;g7,genu of facial nerve;GiA,αpart of gigantocellular retic-ular nucleus;icp,inferior cerebellar peduncle;m7,facial nucleus;mlf,medial longitudinal fasciculus;LPGi,lateral paragigantocellular nucleus;PPy,parapyramidal nucleus;Pr,prepositus nucleus;py,pyra-mid;RMg,raphe magnus;RPa,raphe pallidus nucleus;rs,rubrospinal tract;scp,superior cerebellar peduncle;Sp5,spinal trigeminal nucleus;sp5,spinal trigeminal tract;SSN,superior salivatory nucleus;ts,tec-tospinal tract;VeCb,vestibulocerebellar nucleus;VeL,lateral vestibular nucleus;VeM,medial vestibular nucleus.(Modified,with permission,from Li et al.2010)(209).sublingual glands.A coarse topography exists within the SSN such that neurons projecting to the pterygopalatine ganglion are generally located ventral to those projecting to the sub-mandibular ganglion (65,69,166,261,263,396).Pterygopalatine ganglionIn humans,the pterygopalatine ganglion is located in the pterygopalatine fossa,its shape varies between individuals,being variously described as conical,triangular,or ellip-tical,and it is reported to be approximately 3mm verti-cal,4mm sagittal,and 2mm transversely (377).However,more recently,it has been recognized that the pterygopala-tine ganglion may be bipartite,and a number of associated microganglia may also be present (e.g.,305,324).Within the pterygopalatine ganglion,the preganglionic neurons form cholinergic,nicotinic synapses with postganglionic neurons.Axons of these postganglionic neurons leave the pterygopala-tine ganglion caudally and reach the eye by way of the rami orbitales,the retro-orbital plexus,and the rami oculares (173,316).These postganglionic fibers target a number of ocular structures including the ciliary epithelium and trabec-ular meshwork,as well as blood vessels in the optic nerve,iris,ciliary body,choroid,and episclera (Figs.1and 2).In mam-mals including primates,postganglionic fibers from the ptery-gopalatine ganglion also project to cerebral blood vessels,the lacrimal gland,blood vessels of the nasal mucosa and palate,and the meibomian glands (206,258,314,318,379,401,408).Postganglionic neurons within the pterygopalatine gan-glion are cholinergic,but many also express nNOS and vasoactive intestinal peptide (VIP)(131,173,200,260,305).Also,similar to the ciliary ganglion,some pterygopalatineganglion cells in rats are innervated by collaterals of trigemi-nal afferents (Fig.2)(373).Sympathetic pathways innervating the eye The intermediolateral cell columnThe preganglionic sympathetic neurons that innervate the eye are located in the intermediolateral cell column (IML)in the C8-T2segments of the spinal cord,a region termed the cil-iospinal center of Budge (and Waller).The axons of these preganglionic neurons project to the sympathetic chain gan-glia and travel in the sympathetic trunk to the superior cervical ganglion (148,178).Superior cervical ganglionThe superior cervical ganglion is located at the C1-C3verte-bral level and,in humans,it lies immediately anterior to the common carotid artery bifurcation (422).Within the supe-rior cervical ganglion,preganglionic axons form nicotinic,cholinergic synapses with postganglionic neurons.The axons of these postganglionic neurons project from the superior cer-vical ganglion as the internal carotid nerves to the level of the cavernous sinuses where they form the carotid plexus around the carotid arteries.Many of the fibers destined for the eye form the sympathetic root of the ciliary ganglion,which runs along with the ophthalmic artery to enter the orbit.These fibers then traverse the ciliary ganglion to enter the eye via the short ciliary nerves (281,322).Other sympathetic fibers enter the eye through the long ciliary nerves as well as through the optic canal (322).In addition to containing the neurotransmit-ter noradrenaline,many postganglionic sympathetic neurons projecting to the eye in mammals also express the vasocon-strictor peptide,NPY (41,42,46,123,136,267,305,381),as well as the cotransmitter ATP (294)(Fig.1).Trigeminal nervous system—peripheral reflex controlIn birds and mammals,the nerve branches arising from the trigeminal nerve provide the sensory innervation of the eye.More specifically,the ophthalmic branch of the trigeminal nerve divides into the lacrimal,frontal,and nasociliary nerves.The latter divides into a major branch that travels with the long ciliary nerves to the eye and a minor branch that passes through the ciliary ganglion to enter the eye with the short ciliary nerves (173,198,260,378).Additionally,in monkeys,Ruskell (319)has suggested that there is a small contribution to the sensory innervation of the eye from the maxillary branch of the trigeminal nerve.Within the eye of mammals and birds,sensory fibers are distributed to the:cornea (169,229,362);conjunctiva (83,221);iris including the sphincter muscle,blood vessels,and anterior melanocytes (22,173);ciliary body including theComprehensive Physiology Autonomic Control of the Eyeciliary processes,blood vessels,and a region of the ciliary muscle near the scleral spur(173,321,363);ophthalmic and central retinal arteries(28,35,84,305);choroid including the intrinsic choroidal neurons(336,344,363);episcleral vessels (339,340);and trabecular meshwork(204,338).It has been recognized that the peripheral nerve endings and collaterals of these sensoryfibers contain a number of neuropeptides,and over the past few decades it has been estab-lished that these neuropeptides can also serve as neurotrans-mitters(e.g.,149,173).Thus,it is now clear that the trigemi-nal sensory system can act as a local effector system not only within the various ocular structures that it innervates,but also through collaterals that are reported in both the ciliary and pterygopalatine ganglia(173,260).In general,in mammals including humans,a significant percentage of sensory nerve endings contain SP(173,364,384,385)and CGRP(173,365), and both neuropeptides are present in cells in the trigemi-nal ganglion(122,173,207).In addition,other tachykinins such as neurokinin A,neurokinin B,and neuropeptide K have been reported,as have cholecystokinin(CCK),galanin,and vasopressin(see173for a review).SP,and presumably other neuropeptides,are released from peripheral nerve endings in response to antidromic trigeminal nerve stimulation,as well as peripheral stimulation from capsaicin,bradykinin,histamine, nicotine,and prostaglandins(37,173,227,410,435). Autonomic Control of the PupilThe pupil is essentially an optical element that controls the amount of light striking the retina by acting as an aperture stop:the most light-restrictive element of an optical system. Pupillary diameter,or more precisely iris size,is controlled by two muscles,the sphincter pupillae,which is primarily under the control of the parasympathetic nervous system,and the dilator pupillae,which is primarily under the control of the sympathetic nervous system.Contraction of the sphinc-ter,accompanied by relaxation of the dilator,produces pupil constriction(miosis);while contraction of the dilator,accom-panied by relaxation of the sphincter,produces pupil dilation (mydriasis).These muscles are innervated by postganglionic neurons originating from the ciliary and superior cervical ganglia,which in turn receive preganglionic inputs from the EWpg and IML,respectively(Figs.1and2).Neurons in the EWpg and IML are,in turn,controlled by central nervous system inputs that are influenced by a variety of factors such as retinal irradiance,viewing distance,alertness,and cogni-tive load.The normal human pupil can change diameter from8 to1.5mm,which corresponds to approximately a28-fold change in area.Thus the movement of the iris can account for almost1.5log unit variation in retinal irradiance(281). Although the visual system can operate over a10log unit range of lighting levels through the process of adaptation, it can take several minutes for optimum sensitivity to return after an abrupt increase or decrease in retinal illumination.The rapid control of retinal irradiance by the iris allows the visual system to more quickly regain optimal sensitivity by dampening fast changes in ambient lighting levels and requir-ing less retinal adaptation to a given change in environmental lighting levels(150,426).Changes in pupil size modulate not only retinal illumina-tion,but also depth of focus,optical aberrations,and diffrac-tion.As pupil diameter decreases,depth of focus increases (228,411)and the image degrading effects of optical aber-rations decrease(211,337),but the image degrading effects of diffraction increase(53,58,92).However,over the normal range of pupillary diameter,diffraction impacts image quality less than optical aberrations,and the optimal pupil diameter is therefore approximately between2and4mm(55,167,416). These various factors differentially affect visual performance and,given changing environmental lighting conditions and visual tasks,the nervous system continuously modulates pupil diameter for optimal visual performance.Clearly,a mobile pupil allows the autonomic nervous sys-tem to optimize retinal irradiance,diffraction,ocular aber-rations,and depth of focus despite differing conditions and visual tasks.For example,across a range of daylight(pho-topic)luminance levels,pupil size corresponds to that required for the highest visual acuity(54),and the maximal information capacity of the retinal image(147,205).On the other hand, under low light(scotopic)conditions in which poorer retinal image quality can be tolerated due to the lower resolution of rod photoreceptors,the pupil dilates sufficiently to maximize retinal illumination.Further evidence for the optimization of pupil diameter for differing visual tasks is evident in the pupil-lary near response.When viewing distance changes from far to near,the pupils constrict to increase the depth offield and reduce retinal image defocus(see below for more detail). Overview of the pathways controllingpupil diameterA summary diagram of our current understanding of the affer-ent,central,and efferent pathways controlling pupil diameter are shown in Figure6.Thisfigure shows the iris musculature innervated by autonomic efferents from both the parasympa-thetic and sympathetic components of the autonomic nervous system.Iris musculatureIn a cross section of the iris,the sphincter pupillae can be seen as an annular band of smooth muscle(100-170μm thick;0.7-1.0mm wide)encircling the pupil(Fig.7).The sphincter,which is located in the posterior iris immediately anterior to the pigmented epithelium,interdigitates with the surrounding stroma and connects to dilator musclefibers(see below).The smooth muscle cells of the sphincter are clus-tered in small bundles and connected by gap junctions(45). These gap junctions ensure synchronized contraction of the。
Neural Networks and Evolutionary Computation. Part II Hybrid Approaches in the Neuroscience
Neural Networks and Evolutionary Computation. Part II:Hybrid Approaches in the NeurosciencesGerhard WeißAbstract—This paper series focusses on the intersection of neural networks and evolutionary computation.It is ad-dressed to researchers from artificial intelligence as well as the neurosciences.Part II provides an overview of hybrid work done in the neurosciences,and surveys neuroscientific theories that are bridging the gap between neural and evolutionary computa-tion.According to these theories evolutionary mechanisms like mutation and selection act in real brains in somatic time and are fundamental to learning and developmental processes in biological neural networks.Keywords—Theory of evolutionary learning circuits,theory of selective stabilization of synapses,theory of selective sta-bilization of pre–representations,theory of neuronal group selection.I.IntroductionIn the neurosciences biological neural networks are in-vestigated at different organizational levels,including the molecular level,the level of individual synapses and neu-rons,and the level of whole groups of neurons(e.g.,[11, 36,44]).Several neuroscientific theories have been pro-posed which combine thefields of neural and evolution-ary computation at these different levels.These are the theory of evolutionary learning circuits,the theories of se-lective stabilization of synapses and pre–representations, and the theory of neuronal group selection.According to these theories,neural processes of learning and develop-ment strongly base on evolutionary mechanisms like mu-tation and selection.In other words,according to these theories evolutionary mechanisms play in real brains and nervous systems the same role in somatic time as they do in ecosystems in phylogenetic time.(Other neuroscientific work which is closely related to these theories is described in[28,46,47].)This paper overviews the hybrid work done in the neu-rosciences.Sections II to V survey the four evolutionary theories mentioned above.This includes a description of the major characteristics of these theories as well as a guide to relevant and related literature.Section VI concludes the paper with some general remarks on these theories and their relation to the hybrid approaches proposed in artifi-cial intelligence.The author is with the Institut f¨u r Informatik(H2),Technische Uni-versit¨a t M¨u nchen,D-80290M¨u nchen,Germany.II.The Theory of EvolutionaryLearning CircuitsAccording to the theory of evolutionary learning circuits(TELC for short)neural learning is viewed as the grad-ual modification of the information–processing capabilities of enzymatic neurons through a process of variation andselection in somatic time[12,13].In order to put this more precisely,first a closer look is taken at enzymatic neurons,and then the fundamental claims of the TELCare described.The TELC starts from the point of view that the brainis organized into various types of local networks which contain enzymatic neurons,that is,neurons whosefiringbehavior is controlled by enzymes called excitases.(For details of this control and its underlying biochemical pro-cesses see e.g.[14,15].)These neurons incorporate the principle of double dynamics[15]by operating at two levels of dynamics:at the level of readin or tactilization dynam-ics,the neural input patterns are transduced into chemical–concentration patterns inside the neuron;and at the level of readout dynamics,these chemical patterns are recognized by the excitases.Consequently,the enzymatic neurons themselves are endowed with powerful pattern–recognition capabilities where the excitases are the recognition primi-tives.Both levels of dynamics are gradually deformable as a consequence of the structure–function gradualism(“slight changes in the structure cause slight changes in the func-tion”)in the excitases.As Conrad pointed out,this struc-ture–function gradualism is the key to evolution and evo-lutionary learning in general,and is a important condition for evolutionary adaptability in particular.(Evolutionary adaptability is defined as the extent to which mechanisms of variation and selection can be utilized in order to survive in uncertain and unknown environments[16].)There are three fundamental claims made by the TESC: redundancy of brain tissue,specifity of neurons,and ex-istence of brain–internal selection circuits.According to the claim for redundany,there are many replicas of each type of local network.This means that the brain consists of local networks which are interchangeable in the sense that they are highly similar with respect to their connec-tivity and the properties of their neurons.The claim for specifity says that the excitases are capable of recognizing specific chemical patterns and,with that,cause the enzy-matic neurons tofire in response to specific input patterns. According to the third claim,the brain contains selectioncircuits which direct thefitness–oriented,gradual modifi-cation of the local networks’excitase configurations.These selection circuits include three systems:a testing system which allows to check the consequences(e.g.,pleasure or pain)of the outputs of one or several local networks for the organism;an evaluation system which assignsfitness values to the local networks on the basis of these consequences; and a growth–control system which stimulates or inhibits the production of the nucleic acids which code for the lo-cal networks’excitases on the basis of theirfitness values. The nucleic acids,whose variability is ensured by random somatic recombination and mutation processes,diffuse to neighbouring networks of the same type(where they per-form the same function because of the interchangeability property mentioned above).These claims imply that neu-ral learning proceeds by means of the gradual modification of the excitase configurations in the brain’s local networks through the repeated execution of the following evolution-ary learning cycle:1.Test and evaluation of the enzymatic neuron–basedlocal networks.As a result,afitness value is assigned to each network.2.Selection of local networks.This involves thefitness–oriented regulation of the production of the excitase–coding nucleic acids,as well as their spreading to ad-jacent interchangeable networks.3.Application of somatic recombination and selection tothese nucleic acids.This maintains the range of the excitase configurations.The execution stops when a local network is found which has a sufficiently highfitness.Conrad emphasized that this evolutionary learning cycle is much more efficient than nat-ural evolution because the selection circuits enable an in-tensive selection even if there is hardly a difference between thefitness values of the interchangeable networks. Finally,some references to related work.The TESC is part of extensive work focussing on the differences be-tween the information processing capabilities of biological (molecular)systems and conventional computers;see e.g. [15,16,17].A computational specification of the ESCM which concentrates on the pattern–processing capabilities of the enzymatic neurons,together with its sucessful appli-cation to a robot–control task,is contained in[29,30,31]. Another computational specification which concentrates on the intraneuronal dynamics of enzymatic neurons is de-scribed in[32].A combination of these two specifications is described in[33].Further related work being of particu-lar interest from a computational point of view is presented in[1,18].III.The Theory of Selective Stabilizationof SynapsesThe theory of selective stabilization of synapses(TSSS for short)is presented in[7,8].This theory accounts for neural processes of learning and development by postulat-ing that a somatic,evolutionary selection mechanism acts at the level of synapses and contributes to the wiring pat-tern in the adult brain.Subsequently the neurobiological basis and the major claims of the TSSS are depicted. The neurobiological basis of the TSSS comprises aspects of both neurogenesis and neurogenetics.In vertebrates one can distinguish several processes of brain development. These are the cellular processes of cell division,movement, adhesion,differentiation,and death,and the synaptic pro-cesses of connection formation and elimination.(For de-tails see e.g.[19,20,38].)The TSSS focusses on the“synap-tic aspect”of neurogenesis;it deals with the outgrowth and the stabilization of synapses,and takes the developmental stage where maximal synaptic wiring exists as its initial state.The neurogenetic attidue of the TSSS constitutes a compromise between the preformist(“specified–by–genes”) and the empirist(“specified–by–activity”)view of brain development.It is assumed that the genes involved in brain development,the so–called genetic envelope,only specify the invariant characters of the brain.This includes,in particular,the connections between the main categories of neurons(i.e.,between groups of neurons which are of the same morphological and biochemical type)and the rules of synaptic growth and stabilization.These rules allow for an activity–dependent,epigenetic synapse formation within the neuronal categories.(As Changeux formulated:“The genetic envelope offers a hazily outlined network,the ac-tivity defines its angles.”[3,p.193])The TSSS makes three major claims.First,at the crit-ical stage of maximal connectivity there is a significant but limited redundany within the neuronal categories as regards the specifity of the synapses.Second,at this time of so–called“structural redundany”any synapse may exist under(at least)three states of plasticity:labile,stable,and degenerate.Only the labile and stable synapses transmit nerve impulses,and the acceptable state transitions are those from labile to either stable or degenerate and from stable to labile.Especially,the state transition of a synapse is epigenetically regulated by all signals received by the postsynaptic soma during a given time interval.(The max-imal synaptic connectivity,the mechanisms of its develop-ment,and the regulative and integrative properties of the soma are determinate expressions of the genetic envelope.) Third,the total activity of the developing network leads to the selective stabilization of some synapses,and to the re-gression of their functional equivalents.As a consequence, structural redundancy decreases and neuronal singularity (i.e.,individual connectivity)increases.This provides a plausible explanation of the naturally occuring connection elimination occuring during neural development.For further readings in the TSSS see e.g.[4,5,6,10].IV.The Theory of Selective Stabilizationof Pre–RepresentationsThe theory of selective stabilization of pre–representa-tions(TSSP for short)can be viewed as an extension of the TSSS.This theory provides a selectionist view of neurallearning and development in the adult brain by postulating that somatic selection takes place at the level of neural networks[5,10,27].Similar to the theory of neuronal group selection(section V),the TSSP may be viewed as an attempt to show how neurobiology and psychology are related to each other.There are two major claims made by the TSSP.Thefirst claim is that there exist mental objects or“neural repre-sentations”in the brain.A mental object is defined as a physical state achieved by the correlated and transitory (both electrical and chemical)activity of a cell assembly consisting of a large number of neurons having different singularities.According to the TSSP,three classes of men-tal objects are distinguished.First,primary percepts;these are labile mental objects whose activation depends on the direct interaction with the outside world and is caused by sensory stimulations.Second,stored representations;these are memory objects whose evocation does not demand en-vironmental interaction and whose all–or–none activity be-havior results from a stable,cooperative coupling between the neurons.And third,pre–representations;these are mental objects which are generated before and concomitant with any environmental interaction.Pre–representations are labile and of great variety and variability;they result from the spontaneous but correlatedfiring of neurons or groups of neurons.The second claim made by the TSSP is that learning in the adult brain corresponds to the se-lective stabilization of pre–representations,that means,the transition from selected pre–representations to stored rep-resentations.This requires,in the simplest case,the inter-action with the environment,the criterion of selection is the resonance(i.e.,spatial overlapping orfiring in phase) between a primary percept and a pre–representation.Further literature on the TSSP.In[9]the two selective–stabilization theories,TSSS and TSSP,are embedded in more general considerations on the neural basis of cogni-tion.A formal model of neural learning and development on the basis of the TSSP is described in[22,43].V.The Theory of Neuronal Group SelectionThe theory of neuronal group selection(TNGS for short) or“neural Darwinism”[23,25]is the most rigorous and elaborate hybrid approach in the neurosciences.This the-ory,which has attracted much attention especially in the last few years,bridges the gap between biology and psy-chology by postulating that somatic selection is the key mechanism which establishes the connection between the structure and the function of the brain.As done in the preceding sections,below the major ideas of the TNGS are described.There are three basic claims.First,during prenatal and early postnatal development,primary repertoires of degen-erate neuronal groups were formed epigenetically by selec-tion.According to the TNGS a neuronal group is consid-ered as a local anatomical entity which consists of hundreds to thousands of strongly connected neurons,and degener-ate neuronal groups are groups that have different struc-tures but carry out the same function more or less well (they are nonisomorphic but isofunctional).The concept of degeneracy is fundamental to the TNGS;it implies both structural diversity and functional redundancy and,hence, ensures both a wide range of recognition and the reliabil-ity against the loss of neural tissue.Degeneracy naturally origins from the processes of brain development which are assumed to occur in an epigenetic manner and to elabo-rate several selective events at the cellular level.According to the regulator hypothesis,these complex developmental processes,as well as the selective events accompaning these processes,are guided by a relatively small number of cell adhesion molecules.Second,in the(postnatal)phase of be-havioral experience,a secondary repertoire of functioning neuronal groups is formed by selection among the preexist-ing groups of each primary repertoire.This group selection is accomplished by epigenetic modifications of the synap-tic strenghts without change of the connectivity pattern. According to the dual rules model,these modifications are realized by two synaptic rules that operate upon popula-tions of synapses in a parallel and independent fashion: a presynaptic rule which applies to long–term changes in the whole target neuron and which affects a large num-ber of synapses;and a postsynaptic rule which applies to short–term changes at individual synapses.The function-ing groups are more likely to respond to identical or simi-lar stimuli than the non–selected groups and,hence,con-tribute to the future behavior of the organism.A funda-mental operation of the functional groups is to compete for neurons that belong to other groups;this competition affects the groups’functional properties and is assumed to play a central role in the formation and organization of cerebral cortical maps.Third,reentry–phasic sig-naling over re–entrant(reciprocal and cyclic)connections between different repertoires,in particular between topo-graphic maps–allows for the spatiotemporal correlation of the responses of the repertoires at all levels in the brain. This kind of phasic signaling is viewed as an important mechanism supporting group selection and as being essen-tial both to categorization and the development of con-sciousness.Reentry implies two fundamental neural struc-tures:first,classification couples,that is,re–entrant reper-toires that can perform classifications more complex than a single involved repertoire could do;and second,global map-pings,that is,re-entrant repertoires that correlate sensory input and motor activity.Some brief notes on how the TNGS accounts for psy-chological functions.Following Edelman’s argumentation, categories do not exist apriori in the world(the world is “unlabeled”),and categorization is the fundamental prob-lem facing the nervous system.This problem is solved by means of group selection and reentry.Consequently,cat-egorization largely depends on the organism’s interaction with its environment and turns out to be the central neural operation required for all other operations.Based on this view of categorization,Edelman suggests that memory is “the enhanced ability to categorize or generalize associa-tively,not the storage of features or attributes of objects as a list”[25,p.241]and that learning,in the minimal case,is the“categorization of complexes of adaptive value under conditions of expectancy”[25,p.293].There is a large body of literature on the TNGS.The most detailed depiction of the theory is provided in Edel-man’s book[25].In order to be able to test the TNGS, several computer models have been constructed which em-body the theory’s major ideas.These models are Darwin I[24],Darwin II[26,39,25],and Darwin III[40,41].Re-views of the TNGS can be found in e.g.[21,34,35,42,37].VI.Concluding RemarksThis paper overviewed neuroscientific theories which view real brains as evolutionary systems or“Darwin machines”[2].This point of view is radically opposed to traditional in-structive theories which postulate that brain development is directed epigenetically during an organism’s interaction with its environment by rules for a more or less precise brain wiring.Nowadays most researchers agree that the instructive theories are very likely to be wrong und unre-alistic,and that the evolutionary theories offer interesting and plausible alternatives.In particular,there is an in-creasing number of neurobiological facts and observations described in the literature which indicate that evolutionary mechanisms(and in particular the mechanism of selection) as postulated by the evolutionary theories are indeed fun-damental to the neural processes in our brains.Somefinal notes on the relation between the hybrid work done in the neurosciences and the hybrid work done in artificial intelligence(see part I of this paper series[45]). Whereas the neuroscientific approaches aim at a better un-derstanding of the developmental and learning processes in real brains,the artificial intelligence approaches typically aim at the design of artificial neural networks that are ap-propriate for solving specific real-world tasks.Despite this fundamental difference and its implications,however,there are several aspects and questions which are elementary and significant to both the neuroscientific and the artificial in-telligence approaches:•Symbolic–subsymbolic intersection(e.g.,“What are the neural foundations of high-level,cognitive abili-ties like concept formation?”and“How are symbolic entities encoded in the neural tissue?”),•Brain wiring(e.g.,“What are the principles of neural development?”and“How are the structure and the function of neural networks related to each other?”),•Genetic encoding(e.g.,“How and to what extend are neural networks genetically encoded?”),and •Evolutionary modification(e.g.,“At what network level and at what time scale do evolutionary mechanisms operate?”and“In how far do the evolutionary mech-anisms influence the network structure?”).Because of this correspondence of interests and research topics it would be useful and stimulating for the neurosci-entific and the artificial intelligence community to be aware of each others hybrid work.This requires an increased in-terdisciplinary transparency.To offer such a transparency is a major intention of this paper series.References[1]Akingbehin,K.&Conrad,M.(1989).A hybrid architecture forprogrammable computing and evolutionary learning.Parallel Distributed Computing,6,245–263.[2]Calvin,W.H.(1987).The brain as a Darwin machine.Nature,330,pp.33–43.[3]Changeux,J.–P.(1980).Genetic determinism and epigenesisof the neuronal network:Is there a biological compromise be-tween Chomsky and Piaget?In M.Piatelli–Palmarini(Ed.), Language and learning–The debate between Jean Piaget and Noam Chomsky(pp.184–202).Routledge&Kegan Paul. [4]Changeux,J.–P.(1983).Concluding remarks:On the“singu-larity”of nerve cells and its ontogenesis.In J.–P.Changeux, J.Glowinski,M.Imbert,&F.E.Bloom(Eds.),Molecular and cellular interactions underlying higher brain function(pp.465–478).Elsevier Science Publ.[5]Changeux,J.–P.(1983).L’Homme neuronal.Fayard.[6]Changeux,J.–P.(1985).Remarks on the complexity of thenervous system and its ontogenesis.In J.Mehler&R.Fox (Eds.),Neonate cognition.Beyond the blooming buzzing con-fusion(pp.263–284).Lawrence Erlbaum.[7]Changeux,J.–P.,Courrege,P.,&Danchin,A.(1973).A theoryof the epigenesis of neuronal networks by selective stabilization of synapses.In Proceedings of the National Academy of Sciences USA,70(10),2974–2978.[8]Changeux,J.–P.,&Danchin,A.(1976).Selective stabilizationof developing synapses as a mechanism for the specification of neuronal networks.Nature,264,705–712.[9]Changeux,J.–P.,&Dehaene,S.(1989).Neuronal models ofcognitive functions.Cognition,33,63–109.[10]Changeux,J.–P.,Heidmann,T.,&Patte,P.(1984).Learningby selection.In P.Marler&H.S.Terrace(Eds.),The biology of learning(pp.115–133).Springer.[11]Changeux,J.–P.,&Konoshi,M.(Eds.).(1986).The neural andmolecular basis of learning.Springer.[12]Conrad,M.(1974).Evolutionary learning cicuits.Journal ofTheoretical Biology,46,167–188.[13]Conrad M.(1976).Complementary models of learning andmemory.BioSystems,8,119–138.[14]Conrad,M.(1984).Microscopic–macroscopic interface in bio-logical information processing.BioSystems,16,345–363. [15]Conrad,M.(1985).On design principles for a molecular com-munications of the ACM,28(5),464–480.[16]Conrad,M.(1988).The price of programmability.In R.Herken(Ed.),The universal Turing machine–A half–century survey (pp.285–307).Kammerer&Unverzagt.[17]Conrad,M.(1989).The brain–machine disanalogy.BioSystems,22,197–213.[18]Conrad,M.,Kampfner,R.R.,Kirby,K.G.,Rizki, E.N.,Schleis,G.,Smalz,R.,&Trenary,R.(1989).Towards an ar-tificial brain.BioSystems,23,175–218.[19]Cowan,W.M.(1978).Aspects of neural development.Interna-tional Reviews of Physiology,17,150–91.[20]Cowan,W.M.,Fawcett,J.W.,O’Leary,D.D.M.,&Stanfield,B.B.(1984).Regressive events in neurogenesis.Science,225,1258–1265.[21]Crick,F.(1989).Neural Edelmanism.Trends in Neurosciences,12,240–248.(Reply from G.N.Reeke,R.Michod and F.Crick: Trends in Neurosciences,13,11–14.)[22]Dehaene,S.,Changeux,J.–P.,&Nadal,J.–P.(1987).Neuralnetworks that learn temporal sequences by selection.In Proceed-ings of the National Academy of Sciences USA,84,pp.2727–2731.[23]Edelman,G.M.(1978).Group selection and phasic reentrantsignaling:A theory of higher brain function.In G.M.Edelman &V.B.Mountcastle(Eds.),The mindful brain.Cortical organi-zation and the group–selective theory of higher brain functions (pp.51–100).The MIT Press.[24]Edelman,G.M.(1981).Group selection as the basis for higherbrain function.In F.O.Schmitt,F.G.Worden,G.Adelman&S.G.Dennis(Eds.),The organization of the cerebral cortex (pp.535–563).The MIT Press.[25]Edelman,G.M.(1987).Neural Darwinism.The theory of neu-ronal group selection.Basic Books.[26]Edelman,G.M.,&Reeke,G.N.(1982).Selective networkscapable of representative transformations,limited generaliza-tions,and associative memory.In Proceedings of the National Academy of Sciences USA,79,2091–2095.[27]Heidmann, A.,Heidmann,T.M.,&Changeux,J.–P.(1984).Stabilisation selective de representations neuronales par reso-nance entre“presepresentations”spontanes du reseau cerebral et“percepts”.In C.R.Acad.Sci.Ser.III,299,839–844. [28]Jerne,N.K.(1967).Antibodies and learning:selection vs.in-struction.In G.C.Quarton,T.Melnechuk&F.O.Schmitt (Eds.),The neurosciences:a study program(pp.200–205).Rockefeller University Press.[29]Kampfner,R.R.(1988).Generalization in evolutionary learn-ing with enzymatic neuron–based systems.In M.Kochen&H.M.Hastings(Eds.),Advances in cognitive science.Steps to-ward convergence(pp.190–209).Westview Press,Inc.[30]Kampfner,R.R.,&Conrad,M.(1983).Computational model-ing of evolutionary learning processes in the brain.Bulletin of Mathematical Biology,45(6),931–968.[31]Kampfner,R.R.,&Conrad,M.(1983).Sequential behaviorand stability properties of enzymatic neuron networks.Bulletin of Mathematical Biology,45(6),969–980.[32]Kirby,K.G.,&Conrad,M.(1984).The enzymatic neuron asa reaction–diffusion network of cyclic nucleotides.Bulletin ofMathematical Biology,46,765–782.[33]Kirby,K.G.,&Conrad,M.(1986).Intraneuronal dynamics asa substrate for evolutionary learning.In Physica22D,205–215.[34]Michod,R.E.(1989).Darwinian selection in the brain.Evolu-tion,(3),694–696.[35]Nelson,R.J.(1989).Philosophical issues in Edelman’s neuralDarwinism.Journal of Experimental and Theoretical Artificial Intelligence,1,195–208.[36]Neuroscience Research(1986).Special issue3:Synaptic plastic-ity,memory and learning.[37]Patton,P.,&Parisi,T.(1989).Brains,computation,and selec-tion:An essay review of Gerald Edelman’s Neural Darwinism.Psychobiology,17(3),326–333.[38]Purves,D.,&Lichtman,J.W.(1985).Principles of neural de-velopment.Sinauer Associates Inc.[39]Reeke,G.N.jr.,&Edelman,G.M.(1984).Selective networksand recognition automata.In Annals of the New York Academy of Sciences,426(Special issue on computer culture),181–201.[40]Reeke,G.N.jr.,&Edelman,G.M.(1988).Real brains andartificial intelligence.Deadalus,117(1),143–173.[41]Reeke,G.N.jr.,Sporns,O.,&Edelman,G.M.(1988).Syn-thetic neural modeling:a Darwinian approach to brain theory.In R.Pfeifer,Z.Schreter,F.Fogelman–Soulie&L.Steels(Eds.), Connectionism in perspective.Elsevier.[42]Smoliar,S.(1989).Book review of[25].Artificial Intelligence,39,121–139.[43]Toulouse,G.,Dehaene,S.,&Changeux,J.–P.(1986).Spin glassmodels of learning by selection.In Proceedings of the National Academy of Sciences USA,83,1695–1698.[44]Trends in Neurosciences(1988).Vol.11(4),Special issue:Learn-ing,memory.[45]Weiß,G.(1993,submitted to IEEE World Congress on Compu-tational Intelligence).Neural networks and evolutionary com-putation.Part I:Hybrid approaches in artificial intelligence. [46]Young,J.Z.(1973).Memory as a selective process.In Aus-tralian Academy of Science Report:Symposium on Biological Memory;Australian Academy of Science(pp.25–45).[47]Young,J.Z.(1975).Sources of discovery in neuroscience.InF.G.Worden,J.P.Swazey&G.Edelman(Eds.),The neu-rosciences:Paths of discovery(pp.15–46).Oxford University Press.。
To appear in Computational Neuroscience Trends in Research
Modeling receptivefields withnon-negative sparse codingPatrik O.HoyerNeural Networks Research CentreHelsinki University of TechnologyP.O.Box9800,FIN-02015HUT,Finlandpatrik.hoyer@hut.fiPresented at CNS*2002(Chicago)To appear in:Computational Neuroscience:Trends in Research2003E.De.Schutter(ed),Elsevier,Amsterdam.1IntroductionA fundamental problem in visual neuroscience concerns understanding how the statistics of the environment is reflected in the processing of the early visual system. Research in this area originated almost half a century ago[1]and has recently attracted significant attention as the computational power and the theoretical tools needed to investigate these issues have become available.For a review of thefield, see[14].A large share of this research has concerned the mammalian primary visual cortex, also known as V1.This is largely due to the fact that,ever since the pioneering work of Hubel and Wiesel[6],much is known about the processing performed Preprint submitted to Elsevier Science8August2002there.Neurons in V1can be divided into simple-and complex-cells.Of these,the responses of simple-cells can be reasonably described as a linearfiltering of the visual input.The spatial properties of these linearfilters(called the receptivefields) can be characterized as being localized,oriented,and bandpass[6,3].The question of why this particular type offiltering is performed has undergone considerable debate during the years.In a highly influential paper,Olshausen and Field[11]showed that linear sparse coding(SC)of natural images yielded features qualitatively very similar to recep-tivefields of simple-cells in V1.Subsequently,the very closely related model of in-dependent component analysis(ICA)[7]was shown to give similar results[2].This framework gave the observed receptivefields an information-theoretic explanation which had been lacking in previous work,demonstrating how sensory coding is related to the statistics of the environment.The basic idea in sparse coding is relatively simple:The observed multidimen-sional data(random vector)x is modeled as a linear superposition of some features (called basis vectors)a i,as in x≈∑n i=1a i s i.In the basic image data case,each input vector x consists of the pixel intensities of a small image patch,and the a i likewise contain the pixel intensities of basis patches.The latent variables s i that give the mixing proportions are stochastic and differ for each observed input x.The crucial assumption in the sparse coding framework is that these hidden variables exhibit sparseness.Sparseness is a property independent of scale(variance),and implies that the s i have probability densities which are highly peaked at zero and have heavy tails.Essentially,the idea is that any single typical input vector x can be accurately described using only a few active(significantly non-zero)units s i. However,all of the basis vectors a i are needed to represent the data because the set of active units changes from input to input.The model,as typically used,is completely symmetric around zero.The data is normalized to zero mean,and the sources s i are assumed to have even-symmetric probability densities(i.e.p(s i)=f(|s i|))with a high peak at zero and heavy tails. This also implies that the observed data is completely symmetric with respect to the origin.When this model is learned from natural image data,the learned basis patches have the principal properties of the spatial receptivefields of simple-cells in V1.The neural interpretation of the model is thus that simple-cells in V1per-form sparse coding on the visual input they receive,with the receptivefields closely related to the sparse coding basis vectors and thefiring rates of the neurons repre-senting the latent variables s i.Could the sparse coding model be used more generally to understand how cortical circuits represent their inputs?The fact that areas of the cerebral cortex analyz-ing quite different inputs nevertheless anatomically are relatively similar[13]sug-gests the existence of common computational strategies for cortical sensory cod-ing.Could sparse coding perhaps explain further properties of the visual system,2or possibly properties of other sensory modalities?We believe that to tackle such questions,the model mustfirst be slightly modified,as explained in the next sec-tion.2Non-negative sparse codingThere are at least two obvious ways in which the standard sparse coding image model is unrealistic as a model of V1simple-cell behavior.Perhaps the most glar-ing discrepancy is the fact that in the model each unit s i can,in addition to being effectively silent(close to zero),be either positively or negatively active.This ba-sically means that every feature contributes to representing stimuli of opposing po-larity;for example,the same unit that codes for a dark bar on a bright background also codes for a bright bar on a dark background.This is in clear contrast to the behavior of simple-cells in V1:these neurons tend to have quite low background firing rates and,asfiring rates cannot go negative,thus can only represent one half of the output distribution of a signed feature s i.Another major difference between the model and neural reality is that the input data in the model is double-sided(signed),whereas V1receives the visual data from the lateral geniculate nucleus(LGN)in the form of separated ON-and OFF-channels.Of course,as an abstract model of visual coding,the input data should indeed be(signed)image contrast.But if we on the other hand are interested in how V1recodes its input signals,we must consider separate ON-and OFF-channel input.Thus,if we would like to transform sparse coding from a relatively abstract model of image representation in V1to a concrete model of simple-cell recoding of LGN inputs,the model must be changed.First,our input data should consist of hypo-thetical activities of ON-and OFF-channels in response to natural image patches. Second,all coefficients s i should be restricted to non-negative values.As both the sources s i and the data x thus have non-negative values only,it is logical to assume the same of the features a i.This is due to the fact that if the features contained negative values,this would inevitably be true of the observed data x as well. Although we so far considered non-negativity constraints with the objective of making the model betterfit known neurophysiology,there are also other equally important arguments for non-negativity.In particular,it has been argued that non-negativity can be important for learning parts-based representations[8]:In the stan-dard sparse coding model,the data is described as a combination of elementary features involving both additive and subtractive interactions.The fact that features can‘cancel each other out’using subtraction is contrary to the intuitive notion of combining parts to form a whole.Non-negativity ensures that elementary object features combine additively.3Representations involving only additive combinations of object parts seem espe-cially attractive for modeling higher-level features of images:In[5]it was shown how contour-coding could be learned from the statistics of complex-cell-like re-sponses to natural images.This was possible by representing the model complex-cell responses by a non-negative,sparse,code.Arranging all observed data vectors x into the columns of a matrix X,the corre-sponding unknown sources into the matrix S,and the unknown features a i into the columns of A,the linear model is given by X≈AS.The non-negativity constraints require that all elements of A,and S are zero or positive.The same is assumed of the observed data X.We suggest to use the following objective function for non-negative sparse coding (NNSC)[4,5]:1C(A,S)=Fig.1.Basis vectors from non-negative sparse coding of ON/OFF-channel image data.Left: Generative weights for the ON-channel.Each patch represents the part of one basis vector a i corresponding to the ON-channel.Gray pixels denotes zero weight,brighter pixels are positive weights.Middle:Corresponding weights for the OFF-channel.Right:Weights for ON minus weights for OFF.Fig.2.Basis vectors learned by non-negative matrix factorization.Left:Weights for ON-channel.Middle:Weights for OFF-channel.Right:Weights for ON minus weights for OFF.found at http://www.cis.hut.fi/phoyer/code/.The resulting basis patterns (columns of A)are shown in Figure1.Note how the basis vectors represent Gabor-like image input by elongated ON-and OFF-subregions.This is most clearly seen in the rightmost panel,where the weights from the OFF-channel have been subtracted from those from the ON-channel.These learned features are at least qualitatively very similar to the ones found with the standard sparse coding model applied to symmetric image data.To gain some insight into the roles of the sparseness objective and the non-negativity constraints in these results we can compare the learned decomposition to those given by other methods.For example,in Figure2we show the results of non-negative matrix factorization(NMF)on our data.This is actually a special case of our model,whenλ=0.Note that all features are unoriented;this held for all tested dimensionalities.Thus it seems that sparseness is important for learning oriented, localized features.The table below summarizes ourfindings on the behaviour of different decompositions on the data.localizedyes no yes yesbandpassyes yes no no54Discussion and conclusionsAlthough we here showed only the learning of simple cell-like receptivefields,we believe that this modified sparse coding model can be used to explain higher-order visual receptivefields as well.For example,in[5]it was shown how contour coding and end-stopped receptivefields could be learned by non-negative sparse coding on the simulated responses of complex cells to natural images.It remains to be seen if the proposed model could also be useful in predicting or understanding properties of neurons analyzing other sensory modalities.Finally,we would like to point out one common misconception about models with non-negativity constraints:A frequent objection to non-negativity constraints is the fact that inhibition is widespread in the cortex.However,it is important to note that the constraints are on the activities and the generative representation.In fact, inhibition is a vital part of our model:it is required during inference as an impor-tant part of the competition between units to represent the stimuli as sparsely as possible.AcknowledgementsI wish to thank Aapo Hyv¨a rinen and Jarmo Hurri for useful discussions and helpful comments on an earlier version of the manuscript.References[1]H.B.Barlow.Possible principles underlying the transformations of sensory messages.In W.A.Rosenblith,editor,Sensory Communication,pages217–234.MIT Press, 1961.[2] A.J.Bell and T.J.Sejnowski.The‘independent components’of natural scenes areedgefilters.Vision Research,37:3327–3338,1997.[3]R.L.DeValois,D.G.Albrecht,and L.G.Thorell.Spatial frequency selectivity ofcells in macaque visual cortex.Vision Research,22:545–559,1982.[4]P.O.Hoyer.Non-negative sparse coding.In Neural Networks for Signal ProcessingXII(Proc.IEEE Workshop on Neural Networks for Signal Processing),Martigny, Switzerland,2002.In press.[5]P.O.Hoyer and A.Hyv¨a rinen.A multi-layer sparse coding network learns contourcoding from natural images.Vision Research,42(12):1593–1605,2002.6[6] D.H.Hubel and T.N.Wiesel.Receptivefields and functional architecture of monkeystriate cortex.Journal of Physiology,195:215–243,1968.[7] A.Hyv¨a rinen,J.Karhunen,and E.Oja.Independent Component Analysis.WileyInterscience,2001.[8] D.D.Lee and H.S.Seung.Learning the parts of objects by non-negative matrixfactorization.Nature,401(6755):788–791,1999.[9] D.D.Lee and H.S.Seung.Algorithms for non-negative matrix factorization.InAdvances in Neural Information Processing13(Proc.NIPS*2000).MIT Press,2001.[10]B.A.Olshausen.The Sparsenet software package for MATLAB.Available at/bruno/sparsenet.html.[11]B.A.Olshausen and D.J.Field.Emergence of simple-cell receptivefield propertiesby learning a sparse code for natural images.Nature,381:607–609,1996.[12]P.Paatero and U.Tapper.Positive matrix factorization:A non-negative factor modelwith optimal utilization of error estimates of data values.Environmetrics,5:111–126, 1994.[13]A.J.Rockel,R.W.Hiorns,and T.P.S.Powell.The basic uniformity in structure ofthe neocortex.Brain,103:221–244,1980.[14]E.P.Simoncelli and B. A.Olshausen.Natural image statistics and neuralrepresentation.Annual Review of Neuroscience,24:1193–1215,2001.7。
Brain research
/locate/brainresAvailable online at Research ReportFeatures and timing of the response of single neurons to novelty in the substantia nigra $Charles B.Mikell a ,n ,John P .Sheehy b ,Brett E.Youngerman a ,Robert A.McGovern a ,Teresa J.Wojtasiewicz a ,Andrew K.Chan a ,Seth L.Pullman c ,Qiping Yu c ,Robert R.Goodman d ,Catherine A.Schevon c ,Guy M.McKhann II aaColumbia University Medical Center,Department of Neurological Surgery,710W 168th Street,New York,NY 10032,United States bBarrow Neurological Institute,350West Thomas Road,Phoenix,AZ 85013,United States cColumbia University,Department of Neurology,710W 168th Street,New York,NY 10032,United States dDepartment of Neurosurgery,St.Luke 's-Roosevelt and Beth Israel Hospitals,100010th Ave,New York,NY 10019,United Statesa r t i c l e i n f oArticle history:Accepted 17October 2013Available online 23October 2013Keywords:Deep brain stimulation NoveltyNeurophysiology Human substantia nigra Dopamine neuronsa b s t r a c tSubstantia nigra neurons are known to play a key role in normal cognitive processes and disease states.While animal models and neuroimaging studies link dopamine neurons to novelty detection,this has not been demonstrated electrophysiologically in humans.We used single neuron extracellular recordings in awake human subjects undergoing surgery for Parkinson disease to characterize the features and timing of this response in the substantia nigra.We recorded 49neurons in the substantia ing an auditory oddball task,we showed that they fired more rapidly following novel sounds than repetitive tones.The response was biphasic with peaks at approximately 250ms,comparable to that described in primate studies,and a second peak at 500ms.This response was primarily driven by slower firing neurons as firing rate was inversely correlated to novelty response.Our data provide human validation of the purported role of dopamine neurons in novelty detection and suggest modi fications to proposed models of novelty detection circuitry.&2013The Authors.Published by Elsevier B.V.All rights reserved.1.IntroductionIt is well-established in animal models and neuroimaging studies that substantia nigra (SN)neurons,especially dopa-mine (DA)neurons,avidly respond to reward (Hollerman and Schultz,1998)as well as novelty (Bunzeck and Düzel,2006;Legault and Wise,2001;Li et al.,2003;Ljungberg et al.,1992).A consensus has formed that both increases and decreases in DA neuron firing rate serve as learning signals (Schultz,2007);a large phasic increase in DA from the SN/ventral tegmental area (VTA)to the nucleus accumbens likely strengthens hippocampal inputs when reward is located,leading to memory reinforcement of rewarded behaviors (Goto and Grace,2005).These observations have led to the hypothesis0006-8993/$-see front matter &2013The Authors.Published by Elsevier B.V.All rights reserved./10.1016/j.brainres.2013.10.033$This is an open-access article distributed under the terms of the Creative Commons Attribution License,which permits unrestricted use,distribution,and reproduction in any medium,provided the original author and source arecredited.nCorresponding author.Fax:þ12123052026.E-mail address:cbm2104@ (C.B.Mikell).b r a i n r e s e a r c h 1542(2014)79–84that DA signaling is required for late long-term potentiation (LTP).Some predictions of this model have been confirmed by neuroimaging studies;notably,reward representations in the striatum appear to be enhanced by preceding novel scenes (Guitart-Masip et al.,2010),and memory for visual scenes at 24h after an fMRI experiment correlated to activation within the hippocampus,ventral striatum,and SN/VTA(Bunzeck et al.,2012).However,human electrophysiology studies have not been performed to test key parts of this hypothesis.Specifically,the timing of the substantia nigra response to novelty is unknown;a popular model predicts an early response to novelty,which serves to strengthen hippocampal synapses formed during behaviorally salient novel events (Lisman et al.,2011).Primate studies have also supported a very short latency($100ms)for phasic DA responses to novel stimuli(Ljungberg et al.,1992),but some longer-latency responses have been observed,on the order of 200ms(Mirenowicz and Schultz,1994).However,with some authors positing a“systems-wide computation…[to deter-mine]that there is a high level of novelty or motivational salience”as a requirement for DA release(Lisman et al., 2011),it is critical to know when precisely this release occurs. Using event-related potentials in patients implanted with externalized DBS electrodes,novel stimuli were associated with both an early hippocampal peak($180ms)and a late nucleus accumbens peak($480ms)that was correlated to increased retention in memory(Axmacher et al.,2010). However,this study did not directly address the midbrain. The authors posit that the link between the hippocampal activation and the nucleus accumbens activation is the dopamine system;we used a novelty-oddball task to detect the posited early substantia nigra response that would con-form to this hypothesis,as well as to characterize its features and timing,to a degree not possible with fMRI.This task is also used in electroencephalography(EEG)experiments to evoke the novelty P300,a hippocampally-dependent event related EEG potential(Knight,1996)(Fig.1).2.ResultsForty-nine putative neurons were identified.A total of14959 standard trials,912target trials,and1249novel sound trials were performed.Novel sounds evoked a greater response compared to standard tones over the250–350ms interval (F stimulus¼6.9,p¼0.010(250–300ms)and F stimulus¼7.1, p¼0.009(300–350ms),p¼0.037(both intervals)after correc-tion for multiple comparisons by FDR,Fig.2).A second peak of significantly increasedfiring was seen from500to600ms (F stimulus¼7.1,p¼0.009(500–550ms)and F stimulus¼7.8, p¼0.006(550–600ms),p¼0.037(both intervals)after correc-tion for multiple comparisons by FDR).Responses to target tones and standard tones did not significantly differ over any interval.The aggregated normalizedfiring rate following all standard tone and novel sound trials is displayed in Fig.2,which shows a peak beginning approximately250ms follow-ing novel sounds,and reaching its zenith at300ms.A second peak was seen beginning at500ms.To confirm that substantia nigra neurons differentiated standard and novel tones,we performed principal component analysis(PCA)on the800ms following the stimulus for responses to standard and novel tones.The population of neurons recorded was heterogenous,and not all neurons were expected to discriminate novels and stan-dards.Furthermore,stimulus novelty seems likely to beonly Fig.1–Apparatus and Task.The apparatus consisted of a presentation computer,headphones,a microelectrode,and a neural signal processor.We used a variation on the novelty P300task as described by Fabiani and Friedman (1995)and Knight(1996)to study midbrain responses to stimuli.Novel sounds included environmental sounds (animal noises),mechanical sounds,or non-standard,non‐target musicaltones.Fig.2–Aggregated Neuron Responses.The smoothed normalizedfiring rate was calculated at each millisecond for each trial as described in the text.The curves here depict the averagefiring rate across all novel trials(red)and all standard trials(black).b r a i n r e s e a rc h1542(2014)79–84 80one of many sources of variance within the firing rate of neurons in the SN (other sources could be reward predication error (Hollerman and Schultz,1998)or relevance as a learning signal (Ljungberg et al.,1992),factors that the present experi-ment is not designed to quantify).We therefore used PCA to quantify how discriminatory each neuron was,in terms of stimulus novelty.These analyses were performed in addition to,not instead of,analysis of the original data (Fig.2).As stated in the Section 5,the top three components captured 56.9%of variation across all neurons,and were used for this analysis.Two-sample T -tests were performed comparing principal component scores for novel versus standard stimuli for 49cells over three components (see Section 5);this yielded three tests per neuron,with each neuron having a 14.3%probability of having at least one signi ficant component by chance alone.13Of 49neurons showed at least one compo-nent which signi ficantly differentiated standards and novels (versus 7expected,p ¼0.014by Chi square).We then averagedPCA components of these 13neurons;two peaks were seen,at 300ms and 500ms,similar to the averaged responses (Fig.3).The waveforms and activity of two typical cells of this group are shown in Fig.4.The mean firing rate of this population was 7.93spikes/s,whereas the firing rate of non-discriminatory neurons was 13.6,though this did not meet statistical signi ficance.We therefore investigated how firing rate differentiated response to standard and novel tones.As stated above,the PCA analysis suggested slower-firing cells might respond to novelty more than faster neurons.We thus investigated the fastest firing subset of our population (neurons with firing rates from 18.5to 55.0spikes/s).We thought they might represent GABA-ergic interneurons found within the SN,which have firing rates between 15and 100spikes/s,and are known to have an inhibitory effect on local DA neurons (Tepper et al.,1995).The twelve fastest-firing neurons were suppressed by novel stimuli (Wilcoxon rank-sum,p o 0.05).This effect was no longer signi ficant if we included the next-slowest neuron in the analysis,but it was signi ficant for any faster cutoff selected (e.g.,the fastest eleven,ten,or nine neurons).All neurons in this subgroup demonstrated decreased activity between 250and 350ms following novel stimuli.To further investigate the relation-ship between firing rate and novelty response,we compared firing rate to novelty response across all 49neurons.Overall neuron firing rate was signi ficantly anti-correlated with the mean normalized firing rates during this interval (r ¼À0.34,p ¼0.018,Fig.5),suggesting a slower sub-population of neu-rons was responsible for the overall trend toward increased activity at 300ms following novel stimuli,and faster neurons were responsible for suppression of firing.3.DiscussionThe increased SN neuron activity following novel sounds compared to standard and target tones demonstrates that SN activity does occur in the setting of auditory noveltyinFig.3–Averaged PCA of discriminating neurons.PCA was performed for 49neurons,yielding 13neurons thatdifferentiated standard and novel tones.An Increase in firing rate was noted for novel tones in discriminatory neurons at peaks similar to those described in the aggregateresponse.Fig.4–Example neurons.The normalized firing rates over novel sound trials (red)and standard tone trials (black)are depicted for two individual neurons in the top panels of this figure.These smoothed rates were calculated in the same manner as in Fig.2.The waveforms of the two neurons are depicted in the top-left insets in blue.b r a i n r e s e a rc h 1542(2014)79–8481humans.The neuron firing rate increased 250–350ms after stimulus onset,concurrently with the onset of hippocampally-dependent novelty P300in the cortex,and well after novelty-response hippocampal activity has been shown to begin (o 100ms after stimulus onset)with a similar auditory task in animal models (Ljungberg et al.,1992;Ruusuvirta et al.,1995).Additionally,we found evidence for a later signal from (500–600ms)that might signal behavioral salience back to the hippocampus in accordance with the predictions of the neo-Hebbian model proposed by Lisman et al.,(2011).Further experimentation will be required to verify this possibility.We speculate that slower-firing units that respond to novelty are dopaminergic in character,whereas faster firing units sup-pressed by novelty (420Hz)may be tonically inhibitory inter-neurons within the SN (Grace and Bunney,1980;Grace and Onn,1989;Tepper et al.,1995).Numerous other brain regions are involved in novelty detection,including the prefrontal cortex,thalamus,and primary sensory cortices (Gur et al.,2007).Dorsolateral pre-frontal cortex (DLPFC),in particular,appears to be important for generation of the novelty p3EEG potential (Lovstad et al.,2012).Nucleus accumbens (NAc)is also an important down-stream target of the novelty response;DA responses detected in the shell subregion of the NAc appear to code an integrated signal of novelty and appetitive valence in rodents,and are rapidly habituated with repeated stimulus presentations (Bassareo et al.,2002).These data support the view that the novelty response is important in coding behavioral relevance that mark stimuli as critical for retention in long-term memory (Lisman et al.,2011).This view is consistent with recent findings reported by Zaghloul and colleagues,in which substantia nigra neuron firing re flected unexpected financial rewards,presumably of high behavioral relevance (Zaghloul et al.,2009a ).In all cases,DA likely serves as a signal that thestimulus may have great importance to the organism,and it effects downstream changes consistent with this.Ours is the first report of neuron activation to non-reward-associated novelty in humans.While the primate literature on the subject is almost twenty years old,present models have not fully accounted for the relationship between novelty and reward.One theory suggests that the avidity of DA neuron firing in response to novelty depends heavily upon contextual cues suggesting that novelty is highly predictive of reward (Bunzeck et al.,2012).Latency of cortical activation has been seen to be as early as 85ms when novelty is predictive of reward (Bunzeck et al.,2009).However,this seems like a highly arti ficial situation;in everyday life stimuli likely require extensive evaluation before it is knowable whether they predict reward.In this case the latency seen by Axmacher of $480ms of activation in the nucleus accum-bens seems more plausible (Axmacher et al.,2010);the signal we have seen may represent the posited link between early hippocampal activation and later nucleus accumbens activation.As mentioned in the introduction,we are mindful of literature describing latencies of 100–200ms for onset of DA neuron firing to unconditioned stimuli.The somewhat later onset we describe may be a result of Parkinson pathophysiol-ogy;ERP studies have described lengthening of the onset of the novelty P3potential to 400ms in PD patients (Tsuchiya et al.,2000).Alternately,the larger brains of humans relative to both non-human primates and rodents may perform a signi ficantly more nuanced evaluation of stimuli to deter-mine novelty relative to other mammals;this corresponds to the “systems-wide computation ”required for novelty detec-tion posited by Lisman (Lisman et al.,2011).This computation no doubt involves multiple cortical and subcortical areas.We are mindful that our analysis may have captured some non-DA units in the substantia nigra.This is an intrinsic limitation of extracellular recording in behaving human subjects,given technical and humane constraints on record-ing in the operating room.Additionally,Parkinson disease may have resulted in some loss of DA neurons in this population.However,DA neurons subserving cognitive pro-cesses such as novelty detection are thought to be lost late in the disease process (Gonzalez-Hernandez et al.,2010),and prior reports have described DA-like neurons during DBS surgery (Zaghloul et al.,2009b ).The observed group effects are highly robust,and a number of the units met criteria for DA neurons.Our analysis is fine-grained enough to provide electrophysiological veri fication of neuroimaging findings that suggests a response to novelty in the midbrain,as well as information about the timing of this response.Such electrophysiological investigations into subcortical structures will provide critical information for the development of models of substantia nigra function.4.ConclusionUsing neurophysiological techniques,we have shown that the human substantia nigra encodes novelty.The timecourse of this encoding is biphasic with peaks at 250and 500ms.These data support the view that SN activation isassociatedFig.5–Firing rate versus response to novels.Z-scored firing rate (Y -axis)is compared to firing rate (X -axis).The eight fastest neurons decrease their firing rate in response to novel stimuli,and there is a negative correlation between firing rate and novelty response,suggesting that different populations of neurons respond differently to novelty within the substantia nigra.b r a i n r e s e a rc h 1542(2014)79–8482with both an early phasic signal and a later signal,possibly associated with incorporation of the sensory trace into long-term memory.To understand how SN function contributes to behavior,further experiments will be required to characterize the response properties of this critical structure.5.Experimental procedures5.1.SubjectsPatients with Parkinson disease who were considered candi-dates for surgical therapy were approached for participation in the study.To be considered surgical candidates,patients had to be free of major neurological comorbidities(e.g., Parkinson's-plus)syndromes,including dementia.Ten patients with a mean age of6579.0years and mean disease duration of12.574.1years were enrolled according to a protocol approved by the Columbia University Medical Center Institutional Review Board(CUMC IRB).5.2.Ethics statementAll study procedures were conducted with CUMC IRB approval in accordance with state and federal guidelines.All patients provided informed consent.Following discussion of the study, questions were answered and a copy of the consent form was given to the patient.Consent discussions were documented by study personnel.The CUMC IRB approved all consent and human experimentation procedures in this study.5.3.Microelectrode recordingMicroelectrode recordings were carried out with paired1-m M tungsten-tip electrodes with a power-assisted microdrive.A neural signal processor(Cerberus™from Blackrock Micro-systems,Salt Lake City)recorded from the microelectrode at 30kilosamples/s.The auditory output of the presentation computer was recorded by the neural signal processor to allow for accurate comparison of stimulus times and neural activity.The substantia nigra was identified in accordance with guidelines from Hutchinson et al.,(1998).5.4.Behavioral taskDuring recording of the substantia nigra for clinical purposes, using an auditory oddball methodology similar to that of Knight(1996),subjects were instructed to listen for rarely-repeated“target”tones in a series of repetitive standard tones and non-repeated novel sounds.The pattern of stimuli con-sisted of four to eight repetitive“standard”tones occurring every800ms followed by a“target”tone or a non-target non-repeated“novel”sound(e.g.canine bark;Fig.1).Each stimulus was336ms in length.A pause of random duration between2 and4s followed to jitter the stimuli,and then the pattern was repeated.The subjects were asked to count the target tones. In a version of the task used withfive patients,approximately half of the target tones were replaced by silence,trials which are not considered here.This was done to perform a related experiment;access to the human midbrain is a scarce resource and as many experiments as possible should be done as can be performed safely.Across all recording ses-sions,5.3%of the stimuli were target tones,7.3%of the stimuli were novel sounds,and the rest were standard tones.5.5.Spike sortingThe recorded microelectrode signals were analyzed offline.A clustering algorithm(Wave Clus,Leicester,UK)was used to detect neural spikes and sort them into clusters representing putative neurons(Quiroga et al.,2004).Neurons were included for further analysis if there were at least400spikes recorded and ten novel sound trials performed.Because of the paired electrode tip and spike-sorting,multiple neurons were sometimes recorded simultaneously,and activity was measured for each neuron recorded during each stimulus.5.6.Analysis of responsesTo quantify the response of each neuron to each stimulus(i.e. the response for each trial),the instantaneousfiring rate was calculated for each millisecond by convoluting the neuron spike histogram with a Gaussian kernel(sigma¼25ms). Responses were examined for the800ms following each stimulus.Thisfiring rate curve was then normalized by subtracting the meanfiring rate from the10s prior to stimulus onset and dividing by the standard deviation of the entire recording.We chose to divide by the standard deviation of the entire recording rather than the previous10s to avoid dividing by zero during periods of low neuron activity.We examined the hypothesis that novel stimuli would evoke a greater response from the neurons than repetitive stimuli using a method similar to Zaghloul et al., (2009a).A2-way ANOVA was performed on all of the trial responses from all neurons,with stimulus type(novel vs. standard)as afixed effect,and the recorded neuron as a random effect.This was performed on sixteen50ms bins from0to800ms,and again was repeated for the the case of target vs.standard stimuli.We corrected for multiple com-parisons using the false discovery rate(FDR)method.Most analysis was performed in Matlab(Mathworks,Natick,MA) but FDR was calculated using SAS(SAS Institute,Cary,NC).5.7.Principal component analysis for evaluation of single neuronsBecause clinical considerations limit the number of trials which could be performed while recording any single neuron, we opted to limit the number of comparisons in our exam-ination of single neuron responses by using principal com-ponent ing the princomp command in Matlab,we transformed the800ms response curves following all trials for a given neuron into principal component space.This technique accounts for covaration within a neuron's trial response curves,reorganizing these800dimensional vectors into principal components,a small number of which can account for the vast majority of data variation.Each principal component is an800ms timeseries,and each trial has a score for each component such that the original trial response curve could be reconstructed by summing the products ofb r a i n r e s e a rc h1542(2014)79–8483that trial's scores with their respective po-nents are ranked according to how much data variation they capture;in this case,thefirst component accounted for31.1% of variance,the second for14.5%,the third for11.2%,and the fourth andfifth for7.9%and6.5%,respectively.Given that the top three components accounted for56.9%of the variance, they were retained for further analysis,allowing us to capture the majority of data variation in only three dimensions,and thus reducing multiple comparisons.We therefore limited our analysis to these three components,comparing the scores for novel sounds with those of standard tones by means of two-sample t-test.Principal component analysis and the subsequent t-tests were repeated for every neuron.We then compared the number of discriminatory components we found to what would be expected by chance using a chi square test.Neurons with a component which successfully discrimi-nated between stimulus types were examined further.Author contributionsC.B.M.,C.A.S.,and G.M.M.designed research.C.B.M.,J.P.S., B.E.Y.,R.A.M.,T.J.W., A.K.C.,S.L.P.,Q.Y.,R.R.G.,performed research. C.B.M.,J.P.S., C.A.S.,and G.M.M.analyzed data.C.B.M.,J.P.S.,A.K.C.,and G.M.M.wrote the paper.AcknowledgmentsDr.Charles Mikell is funded by a Janssen Translational Neurosciences Fellowship.r e f e r e n c e sAxmacher,N.,Cohen,M.X.,Fell,J.,Haupt,S.,Dümpelmann,M., Elger,C.E.,Schlaepfer,T.E.,Lenartz,D.,Sturm,V.,Ranganath,C.,2010.Intracranial EEG correlates of expectancy andmemory formation in the human hippocampus and nucleus accumbens.Neuron65,541–549.Bassareo,V.,De Luca,M.A.,Di Chiara,G.,2002.Differential expression of motivational stimulus properties by dopamine in nucleus accumbens shell versus core and prefrontal cortex.J.Neurosci.22,4709–4719.Bunzeck,N.,Düzel,E.,2006.Absolute coding of stimulus novelty in the human substantia nigra/VTA.Neuron51,369–379. Bunzeck,N.,Doeller,C.F.,Fuentemilla,L.,Dolan,R.J.,Düzel,E., 2009.Reward motivation accelerates the onset of neuralnovelty signals in humans to85ms.Curr.Biol.19,1294–1300. Bunzeck,N.,Doeller,C.F.,Dolan,R.J.,Düzel,E.,2012.Contextual interaction between novelty and reward processing within the mesolimbic system.Hum.Brain Mapp.33,1309–1324. Fabiani,M.,Friedman,D.,1995.Changes in brain activity patterns in aging:the novelty oddball.Psychophysiology32,579–594. Gonzalez-Hernandez,T.,Cruz-Muros,I.,Afonso-Oramas,D., Salas-Hernandez,J.,Castro-Hernandez,J.,2010.Vulnerability of mesostriatal dopaminergic neurons in Parkinson's disease.Front.Neuroanat.4,140.Goto,Y.,Grace,A.A.,2005.Dopaminergic modulation of limbic and cortical drive of nucleus accumbens in goal-directedbehavior.Nat.Neurosci.8,805–812.Grace,A.A.,Bunney,B.S.,1980.Nigral dopamine neurons: intracellular recording and identification with L-dopainjection and histofluorescence.Science210,654–656.Grace,A.A.,Onn,S.P.,1989.Morphology and electrophysiological properties of immunocytochemically identified rat dopamine neurons recorded in vitro.J.Neurosci.9,3463–3481.Guitart-Masip,M.,Bunzeck,N.,Stephan,K.E.,Dolan,R.J.,Düzel,E.,2010.Contextual novelty changes reward representationsin the striatum.J.Neurosci.30,1721–1726.Gur,R.C.,Turetsky,B.I.,Loughead,J.,Waxman,J.,Snyder,W., Ragland,J.D.,Elliott,M.A.,Bilker,W.B.,Arnold,S.E.,Gur,R.E., 2007.Hemodynamic responses in neural circuitries fordetection of visual target and novelty:an event-related fMRI study.Hum.Brain Mapp.28,263–274.Hollerman,J.R.,Schultz,W.,1998.Dopamine neurons report an error in the temporal prediction of reward during learning.Nat.Neurosci.1,304–309.Hutchison,W.D.,Allan,R.J.,Opitz,H.,Levy,R.,Dostrovsky,J.O., Lang,A.E.,Lozano,A.M.,1998.Neurophysiologicalidentification of the subthalamic nucleus in surgery forParkinson's disease.Ann.Neurol.44,622–628.Knight,R.,1996.Contribution of human hippocampal region to novelty detection.Nature383,256–259.Legault,M.,Wise,R.A.,2001.Novelty-evoked elevations of nucleus accumbens dopamine:dependence on impulseflow from the ventral subiculum and glutamatergicneurotransmission in the ventral tegmental area.Eur.J.Neurosci.13,819–828.Li,S.,Cullen,W.K.,Anwyl,R.,Rowan,M.J.,2003.Dopamine-dependent facilitation of LTP induction in hippocampal CA1 by exposure to spatial novelty.Nat.Neurosci.6,526–531. Lisman,J.,Grace,A.A.,Düzel,E.,2011.A neoHebbian framework for episodic memory;role of dopamine-dependent late LTP.Trends Neurosci.34,536–547.Ljungberg,T.,Apicella,P.,Schultz,W.,1992.Responses of monkey dopamine neurons during learning of behavioral reactions.J.Neurophysiol.67,145–163.Lovstad,M.,Funderud,I.,Lindgren,M.,Endestad,T.,Due-Tonnessen,P.,Meling,T.,Voytek,B.,Knight,R.T.,Solbakk,A.K.,2012.Contribution of subregions of human frontal cortexto novelty processing.J.Cognitive Neurosci.24,378–395. Mirenowicz,J.,Schultz,W.,1994.Importance of unpredictability for reward responses in primate dopamine neurons.J.Neurophysiol.72,1024–1027.Quiroga,R.Q.,Nadasdy,Z.,Ben-Shaul,Y.,2004.Unsupervised spike detection and sorting with wavelets andsuperparamagnetic clustering.Neural Comput.16,1661–1687. Ruusuvirta,T.,Korhonen,T.,Penttonen,M.,Arikoski,J.,Kivirikko, K.,1995.Behavioral and hippocampal evoked responses in an auditory oddball situation when an unconditioned stimulus is paired with deviant tones in the cat:experiment II.Int.J.Psychophysiol.20,41–47.Schultz,W.,2007.Multiple dopamine functions at different time courses.Annu.Rev.Neurosci.30,259–288.Tepper,J.M.,Martin,L.P.,Anderson,D.R.,1995.GABAA receptor-mediated inhibition of rat substantia nigra dopaminergicneurons by pars reticulata projection neurons.J.Neurosci.15, 3092–3103.Tsuchiya,H.,Yamaguchi,S.,Kobayashi,S.,2000.Impaired novelty detection and frontal lobe dysfunction in Parkinson's disease.Neuropsychologia38,645–654.Zaghloul,K.A.,Blanco,J.A.,Weidemann,C.T.,McGill,K.,Jaggi,J.L., Baltuch,G.H.,Kahana,M.J.,2009a.Human substantia nigra neurons encode unexpectedfinancial rewards.Science323, 1496–1499.Zaghloul,K.A.,Blanco,J.A.,Weidemann,C.T.,McGill,K.,Jaggi,J.L., Baltuch,G.H.,Kahana,M.J.,2009b.Human substantia nigra neurons encode unexpectedfinancial rewards.Science323, 1496–1499.b r a i n r e s e a rc h1542(2014)79–84 84。
英文阅读:猿和人类大脑的进化历程被改写
A new study published in the Cell Pressjournal Current Biology on October 2 couldrewrite the story of ape and human brain evolution. While the neocortex of the brain has been called "the crowning achievement ofevolution and the biological substrateof human mental prowess," newly reported evolutionary1 rate comparisons show that the cerebellum expanded up to six times faster than anticipated throughout the evolution of apes, including humans. The findings suggest that technical intelligence was likely at least as important as social intelligence in human cognitive2 evolution, the researchers say."Our results highlight a previously3 unappreciated role of the cerebellum in ape and human brain evolution that has the potential to refocus researchers' thinking about how and why the brains in these species have become distinct and to shift attention away from an almost exclusive focus on the neocortex as the seat of our humanity," says Robert Barton of Durham University in the United Kingdom.The cerebellum had been seen primarily as a brain region involvedin movement control, adds Chris Venditti of the University of Reading. But more recent evidence has begun to suggest that the cerebellum has a broader range of functions. The cerebellum also contains anintriguingly4 large number of densely5 packed neurons."In humans, the cerebellum contains about 70 billion neurons --four times more than in the neocortex," Barton says. "Nobody reallyknows what all these neurons are for, but they must be doing something important."The neocortex had gotten most of the attention in part because itis such a large structure to begin with. As a result, in looking at variation in the size of various brain regions, the neocortex appeared to show the most expansion. But much of that increase in size could be explained away by the size of the animal as a whole. Sperm6 whales have a neocortex that is proportionally larger than that of humans, for example.By using a comparative method that controlled for those differences in the way the two brain structures correlate, Barton and Venditti uncovered a striking pattern: both nonhuman apes and humans depart from the otherwise tight correlation7 in size between the cerebellum and neocortex found across other primates8 due to relatively9 rapid evolutionary expansion of the cerebellum.Barton and Venditti say that the cerebellum seems to beparticularly involved in the temporal organization of complex behavioral sequences, such as those involved in making and using tools, for instance. Interestingly, evidence is now emerging for a critical role of the cerebellum in language, too.While plenty of work remains10, the new study establishes the cerebellum as "a new frontier for investigations11 into the neural12 basis of advanced cognitive abilities," the researchers say.词汇解析:1 evolutionaryadj.进化的;演化的,演变的;[生]进化论的参考例句:Life has its own evolutionary process.生命有其自身的进化过程。
Deep Sparse Rectifier Neural Networks
Deep Sparse Rectifier Neural NetworksXavier Glorot Antoine Bordes Yoshua BengioDIRO,Universit´e de Montr´e al Montr´e al,QC,Canada glorotxa@iro.umontreal.ca Heudiasyc,UMR CNRS6599UTC,Compi`e gne,FranceandDIRO,Universit´e de Montr´e alMontr´e al,QC,Canadaantoine.bordes@hds.utc.frDIRO,Universit´e de Montr´e alMontr´e al,QC,Canadabengioy@iro.umontreal.caAbstractWhile logistic sigmoid neurons are more bi-ologically plausible than hyperbolic tangentneurons,the latter work better for train-ing multi-layer neural networks.This pa-per shows that rectifying neurons are aneven better model of biological neurons andyield equal or better performance than hy-perbolic tangent networks in spite of thehard non-linearity and non-differentiabilityat zero,creating sparse representations withtrue zeros,which seem remarkably suitablefor naturally sparse data.Even though theycan take advantage of semi-supervised setupswith extra-unlabeled data,deep rectifier net-works can reach their best performance with-out requiring any unsupervised pre-trainingon purely supervised tasks with large labeleddatasets.Hence,these results can be seen asa new milestone in the attempts at under-standing the difficulty in training deep butpurely supervised neural networks,and clos-ing the performance gap between neural net-works learnt with and without unsupervisedpre-training.1IntroductionMany differences exist between the neural network models used by machine learning researchers and those used by computational neuroscientists.This is in part Appearing in Proceedings of the14th International Con-ference on Artificial Intelligence and Statistics(AISTATS) 2011,Fort Lauderdale,FL,USA.Volume15of JMLR: W&CP15.Copyright2011by the authors.because the objective of the former is to obtain com-putationally efficient learners,that generalize well to new examples,whereas the objective of the latter is to abstract out neuroscientific data while obtaining ex-planations of the principles involved,providing predic-tions and guidance for future biological experiments. Areas where both objectives coincide are therefore particularly worthy of investigation,pointing towards computationally motivated principles of operation in the brain that can also enhance research in artificial intelligence.In this paper we show that two com-mon gaps between computational neuroscience models and machine learning neural network models can be bridged by using the following linear by part activa-tion:max(0,x),called the rectifier(or hinge)activa-tion function.Experimental results will show engaging training behavior of this activation function,especially for deep architectures(see Bengio(2009)for a review), i.e.,where the number of hidden layers in the neural network is3or more.Recent theoretical and empirical work in statistical machine learning has demonstrated the importance of learning algorithms for deep architectures.This is in part inspired by observations of the mammalian vi-sual cortex,which consists of a chain of processing elements,each of which is associated with a different representation of the raw visual input.This is partic-ularly clear in the primate visual system(Serre et al., 2007),with its sequence of processing stages:detection of edges,primitive shapes,and moving up to gradu-ally more complex visual shapes.Interestingly,it was found that the features learned in deep architectures resemble those observed in thefirst two of these stages (in areas V1and V2of visual cortex)(Lee et al.,2008), and that they become increasingly invariant to factors of variation(such as camera movement)in higher lay-ers(Goodfellow et al.,2009).Deep Sparse Rectifier Neural NetworksRegarding the training of deep networks,something that can be considered a breakthrough happened in2006,with the introduction of Deep Belief Net-works(Hinton et al.,2006),and more generally the idea of initializing each layer by unsupervised learn-ing(Bengio et al.,2007;Ranzato et al.,2007).Some authors have tried to understand why this unsuper-vised procedure helps(Erhan et al.,2010)while oth-ers investigated why the original training procedure for deep neural networks failed(Bengio and Glorot,2010). From the machine learning point of view,this paper brings additional results in these lines of investigation. We propose to explore the use of rectifying non-linearities as alternatives to the hyperbolic tangent or sigmoid in deep artificial neural networks,in ad-dition to using an L1regularizer on the activation val-ues to promote sparsity and prevent potential numer-ical problems with unbounded activation.Nair and Hinton(2010)present promising results of the influ-ence of such units in the context of Restricted Boltz-mann Machines compared to logistic sigmoid activa-tions on image classification tasks.Our work extends this for the case of pre-training using denoising auto-encoders(Vincent et al.,2008)and provides an exten-sive empirical comparison of the rectifying activation function against the hyperbolic tangent on image clas-sification benchmarks as well as an original derivation for the text application of sentiment analysis.Our experiments on image and text data indicate that training proceeds better when the artificial neurons are either offor operating mostly in a linear regime.Sur-prisingly,rectifying activation allows deep networks to achieve their best performance without unsupervised pre-training.Hence,our work proposes a new contri-bution to the trend of understanding and merging the performance gap between deep networks learnt with and without unsupervised pre-training(Erhan et al., 2010;Bengio and Glorot,2010).Still,rectifier net-works can benefit from unsupervised pre-training in the context of semi-supervised learning where large amounts of unlabeled data are provided.Furthermore, as rectifier units naturally lead to sparse networks and are closer to biological neurons’responses in their main operating regime,this work also bridges(in part)a machine learning/neuroscience gap in terms of acti-vation function and sparsity.This paper is organized as follows.Section2presents some neuroscience and machine learning background which inspired this work.Section3introduces recti-fier neurons and explains their potential benefits and drawbacks in deep networks.Then we propose an experimental study with empirical results on image recognition in Section4.1and sentiment analysis in Section4.2.Section5presents our conclusions.2Background2.1Neuroscience ObservationsFor models of biological neurons,the activation func-tion is the expectedfiring rate as a function of the total input currently arising out of incoming signals at synapses(Dayan and Abott,2001).An activation function is termed,respectively antisymmetric or sym-metric when its response to the opposite of a strongly excitatory input pattern is respectively a strongly in-hibitory or excitatory one,and one-sided when this response is zero.The main gaps that we wish to con-sider between computational neuroscience models and machine learning models are the following:•Studies on brain energy expense suggest that neurons encode information in a sparse and dis-tributed way(Attwell and Laughlin,2001),esti-mating the percentage of neurons active at the same time to be between1and4%(Lennie,2003).This corresponds to a trade-offbetween richness of representation and small action potential en-ergy expenditure.Without additional regulariza-tion,such as an L1penalty,ordinary feedforward neural nets do not have this property.For ex-ample,the sigmoid activation has a steady state regime around12,therefore,after initializing with small weights,all neuronsfire at half their satura-tion regime.This is biologically implausible and hurts gradient-based optimization(LeCun et al., 1998;Bengio and Glorot,2010).•Important divergences between biological and machine learning models concern non-linear activation functions.A common biological model of neuron,the leaky integrate-and-fire(or LIF)(Dayan and Abott,2001),gives the follow-ing relation between thefiring rate and the input current,illustrated in Figure1(left):f(I)=τlogE+RI−V rE+RI−V th+t ref−1,if E+RI>V th0,if E+RI≤V thwhere t ref is the refractory period(minimal time between two action potentials),I the input cur-rent,V r the resting potential and V th the thresh-old potential(with V th>V r),and R,E,τthe membrane resistance,potential and time con-stant.The most commonly used activation func-tions in the deep learning and neural networks lit-erature are the standard logistic sigmoid and theXavier Glorot,Antoine Bordes,YoshuaBengioFigure1:Left:Common neural activation function motivated by biological data.Right:Commonly used activation functions in neural networks literature:logistic sigmoid and hyperbolic tangent(tanh).hyperbolic tangent(see Figure1,right),which areequivalent up to a linear transformation.The hy-perbolic tangent has a steady state at0,and istherefore preferred from the optimization stand-point(LeCun et al.,1998;Bengio and Glorot,2010),but it forces an antisymmetry around0which is absent in biological neurons.2.2Advantages of SparsitySparsity has become a concept of interest,not only incomputational neuroscience and machine learning butalso in statistics and signal processing(Candes andTao,2005).It wasfirst introduced in computationalneuroscience in the context of sparse coding in the vi-sual system(Olshausen and Field,1997).It has beena key element of deep convolutional networks exploit-ing a variant of auto-encoders(Ranzato et al.,2007,2008;Mairal et al.,2009)with a sparse distributedrepresentation,and has also become a key ingredientin Deep Belief Networks(Lee et al.,2008).A sparsitypenalty has been used in several computational neuro-science(Olshausen and Field,1997;Doi et al.,2006)and machine learning models(Lee et al.,2007;Mairalet al.,2009),in particular for deep architectures(Leeet al.,2008;Ranzato et al.,2007,2008).However,inthe latter,the neurons end up taking small but non-zero activation orfiring probability.We show here thatusing a rectifying non-linearity gives rise to real zerosof activations and thus truly sparse representations.From a computational point of view,such representa-tions are appealing for the following reasons:•Information disentangling.One of theclaimed objectives of deep learning algo-rithms(Bengio,2009)is to disentangle thefactors explaining the variations in the data.Adense representation is highly entangled becausealmost any change in the input modifies most ofthe entries in the representation vector.Instead,if a representation is both sparse and robust tosmall input changes,the set of non-zero featuresis almost always roughly conserved by smallchanges of the input.•Efficient variable-size representation.Dif-ferent inputs may contain different amounts of in-formation and would be more conveniently repre-sented using a variable-size data-structure,whichis common in computer representations of infor-mation.Varying the number of active neuronsallows a model to control the effective dimension-ality of the representation for a given input andthe required precision.•Linear separability.Sparse representations arealso more likely to be linearly separable,or moreeasily separable with less non-linear machinery,simply because the information is represented ina high-dimensional space.Besides,this can reflectthe original data format.In text-related applica-tions for instance,the original raw data is alreadyvery sparse(see Section4.2).•Distributed but sparse.Dense distributed rep-resentations are the richest representations,be-ing potentially exponentially more efficient thanpurely local ones(Bengio,2009).Sparse repre-sentations’efficiency is still exponentially greater,with the power of the exponent being the numberof non-zero features.They may represent a goodtrade-offwith respect to the above criteria.Nevertheless,forcing too much sparsity may hurt pre-dictive performance for an equal number of neurons,because it reduces the effective capacity of the model.Deep Sparse Rectifier NeuralNetworksFigure 2:Left:Sparse propagation of activations and gradients in a network of rectifier units.The input selects a subset of active neurons and computation is linear in this subset.Right:Rectifier and softplus activation functions.The second one is a smooth version of the first.3Deep Rectifier Networks3.1Rectifier NeuronsThe neuroscience literature (Bush and Sejnowski,1995;Douglas and al.,2003)indicates that corti-cal neurons are rarely in their maximum saturation regime ,and suggests that their activation function can be approximated by a rectifier.Most previous stud-ies of neural networks involving a rectifying activation function concern recurrent networks (Salinas and Ab-bott,1996;Hahnloser,1998).The rectifier function rectifier(x )=max(0,x )is one-sided and therefore does not enforce a sign symmetry 1or antisymmetry 1:instead,the response to the oppo-site of an excitatory input pattern is 0(no response).However,we can obtain symmetry or antisymmetry by combining two rectifier units sharing parameters.Advantages The rectifier activation function allows a network to easily obtain sparse representations.For example,after uniform initialization of the weights,around 50%of hidden units continuous output val-ues are real zeros,and this fraction can easily increase with sparsity-inducing regularization.Apart from be-ing more biologically plausible,sparsity also leads to mathematical advantages (see previous section).As illustrated in Figure 2(left),the only non-linearity in the network comes from the path selection associ-ated with individual neurons being active or not.For a given input only a subset of neurons are active .Com-putation is linear on this subset:once this subset of neurons is selected,the output is a linear function of1The hyperbolic tangent absolute value non-linearity |tanh(x )|used by Jarrett et al.(2009)enforces sign symme-try.A tanh(x )non-linearity enforces sign antisymmetry.the input (although a large enough change can trigger a discrete change of the active set of neurons).The function computed by each neuron or by the network output in terms of the network input is thus linear by parts.We can see the model as an exponential num-ber of linear models that share parameters (Nair and Hinton,2010).Because of this linearity,gradients flow well on the active paths of neurons (there is no gra-dient vanishing effect due to activation non-linearities of sigmoid or tanh units),and mathematical investi-gation is putations are also cheaper:there is no need for computing the exponential function in activations,and sparsity can be exploited.Potential Problems One may hypothesize that the hard saturation at 0may hurt optimization by block-ing gradient back-propagation.To evaluate the poten-tial impact of this effect we also investigate the soft-plus activation:softplus (x )=log (1+e x )(Dugas et al.,2001),a smooth version of the rectifying non-linearity.We lose the exact sparsity,but may hope to gain eas-ier training.However,experimental results (see Sec-tion 4.1)tend to contradict that hypothesis,suggesting that hard zeros can actually help supervised training.We hypothesize that the hard non-linearities do not hurt so long as the gradient can propagate along some paths ,i.e.,that some of the hidden units in each layer are non-zero.With the credit and blame assigned to these ON units rather than distributed more evenly,we hypothesize that optimization is easier.Another prob-lem could arise due to the unbounded behavior of the activations;one may thus want to use a regularizer to prevent potential numerical problems.Therefore,we use the L 1penalty on the activation values,which also promotes additional sparsity.Also recall that,in or-der to efficiently represent symmetric/antisymmetric behavior in the data,a rectifier network would needXavier Glorot,Antoine Bordes,Yoshua Bengiotwice as many hidden units as a network of symmet-ric/antisymmetric activation functions.Finally,rectifier networks are subject to ill-conditioning of the parametrization.Biases and weights can be scaled in different (and consistent)ways while preserving the same overall network function.More precisely,consider for each layer of depth i of the network a scalar αi ,and scaling the parameters asW i =W iαi and b i =b i ij =1αj.The output units values then change as follow:s =sn j =1αj .Therefore,aslong as nj =1αj is 1,the network function is identical.3.2Unsupervised Pre-trainingThis paper is particularly inspired by the sparse repre-sentations learned in the context of auto-encoder vari-ants,as they have been found to be very useful intraining deep architectures (Bengio,2009),especially for unsupervised pre-training of neural networks (Er-han et al.,2010).Nonetheless,certain difficulties arise when one wants to introduce rectifier activations into stacked denois-ing auto-encoders (Vincent et al.,2008).First,the hard saturation below the threshold of the rectifier function is not suited for the reconstruction units.In-deed,whenever the network happens to reconstruct a zero in place of a non-zero target,the reconstruc-tion unit can not backpropagate any gradient.2Sec-ond,the unbounded behavior of the rectifier activation also needs to be taken into account.In the follow-ing,we denote ˜x the corrupted version of the input x ,σ()the logistic sigmoid function and θthe model pa-rameters (W enc ,b enc ,W dec ,b dec ),and define the linear recontruction function as:f (x,θ)=W dec max(W enc x +b enc ,0)+b dec .Here are the several strategies we have experimented:e a softplus activation function for the recon-struction layer,along with a quadratic cost:L (x,θ)=||x −log(1+exp(f (˜x ,θ)))||2.2.Scale the rectifier activation values coming from the previous encoding layer to bound them be-tween 0and 1,then use a sigmoid activation func-tion for the reconstruction layer,along with a cross-entropy reconstruction cost.L (x,θ)=−x log(σ(f (˜x ,θ)))−(1−x )log(1−σ(f (˜x ,θ))).2Why is this not a problem for hidden layers too?we hy-pothesize that it is because gradients can still flow throughthe active (non-zero),possibly helping rather than hurting the assignment of credit.e a linear activation function for the reconstruc-tion layer,along with a quadratic cost.We triedto use input unit values either before or after the rectifier non-linearity as reconstruction targets.(For the first layer,raw inputs are directly used.)e a rectifier activation function for the recon-struction layer,along with a quadratic cost.The first strategy has proven to yield better gener-alization on image data and the second one on text data.Consequently,the following experimental study presents results using those two.4Experimental StudyThis section discusses our empirical evaluation of recti-fier units for deep networks.We first compare them to hyperbolic tangent and softplus activations on image benchmarks with and without pre-training,and then apply them to the text task of sentiment analysis.4.1Image RecognitionExperimental setup We considered the image datasets detailed below.Each of them has a train-ing set (for tuning parameters),a validation set (for tuning hyper-parameters)and a test set (for report-ing generalization performance).They are presented according to their number of training/validation/test examples,their respective image sizes,as well as their number of classes:•MNIST (LeCun et al.,1998):50k/10k/10k,28×28digit images,10classes.•CIFAR10(Krizhevsky and Hinton,2009):50k/5k/5k,32×32×3RGB images,10classes.•NISTP:81,920k/80k/20k,32×32character im-ages from the NIST database 19,with randomized distortions (Bengio and al,2010),62classes.This dataset is much larger and more difficult than the original NIST (Grother,1995).•NORB:233,172/58,428/58,320,taken from Jittered-Cluttered NORB (LeCun et al.,2004).Stereo-pair images of toys on a cluttered background,6classes.The data has been prepro-cessed similarly to (Nair and Hinton,2010):we subsampled the original 2×108×108stereo-pair images to 2×32×32and scaled linearly the image in the range [−1,1].We followed the procedure used by Nair and Hinton (2010)to create the validation set.Deep Sparse Rectifier Neural NetworksTable1:Test error on networks of depth3.Bold results represent statistical equivalence between similar ex-periments,with and without pre-training,under the null hypothesis of the pairwise test with p=0.05.Neuron MNIST CIF AR10NISTP NORB With unsupervised pre-trainingRectifier 1.20%49.96%32.86%16.46% Tanh 1.16%50.79%35.89%17.66% Softplus 1.17%49.52%33.27%19.19% Without unsupervised pre-trainingRectifier 1.43%50.86%32.64%16.40% Tanh 1.57%52.62%36.46%19.29% Softplus 1.77%53.20%35.48%17.68% For all experiments except on the NORB data(Le-Cun et al.,2004),the models we used are stacked denoising auto-encoders(Vincent et al.,2008)with three hidden layers and1000units per layer.The ar-chitecture of Nair and Hinton(2010)has been used on NORB:two hidden layers with respectively4000 and2000units.We used a cross-entropy reconstruc-tion cost for tanh networks and a quadratic cost over a softplus reconstruction layer for the rectifier and softplus networks.We chose masking noise as the corruption process:each pixel has a probability of0.25of being artificially set to0.The unsuper-vised learning rate is constant,and the following val-ues have been explored:{.1,.01,.001,.0001}.We se-lect the model with the lowest reconstruction error. For the supervisedfine-tuning we chose a constant learning rate in the same range as the unsupervised learning rate with respect to the supervised valida-tion error.The training cost is the negative log likeli-hood−log P(correct class|input)where the probabil-ities are obtained from the output layer(which imple-ments a softmax logistic regression).We used stochas-tic gradient descent with mini-batches of size10for both unsupervised and supervised training phases.To take into account the potential problem of rectifier units not being symmetric around0,we use a vari-ant of the activation function for whichhalf of the units output values are multiplied by-1.This serves to cancel out the mean activation value for each layer and can be interpreted either as inhibitory neurons or simply as a way to equalize activations numerically. Additionally,an L1penalty on the activations with a coefficient of0.001was added to the cost function dur-ing pre-training andfine-tuning in order to increase the amount of sparsity in the learned representations. Main results Table1summarizes the results on networks of3hidden layers of1000hidden units each,Figure3:Influence offinal sparsity on accu-racy.200randomly initialized deep rectifier networks were trained on MNIST with various L1penalties(from 0to0.01)to obtain different sparsity levels.Results show that enforcing sparsity of the activation does not hurtfinal performance until around85%of true zeros.comparing all the neuron types3on all the datasets, with or without unsupervised pre-training.In the lat-ter case,the supervised training phase has been carried out using the same experimental setup as the one de-scribed above forfine-tuning.The main observations we make are the following:•Despite the hard threshold at0,networks trained with the rectifier activation function canfind lo-cal minima of greater or equal quality than those obtained with its smooth counterpart,the soft-plus.On NORB,we tested a rescaled version of the softplus defined by1αsoftplus(αx),which allows to interpolate in a smooth manner be-tween the softplus(α=1)and the rectifier(α=∞).We obtained the followingα/test error cou-ples:1/17.68%,1.3/17.53%,2/16.9%,3/16.66%, 6/16.54%,∞/16.40%.There is no trade-offbe-tween those activation functions.Rectifiers are not only biologically plausible,they are also com-putationally efficient.•There is almost no improvement when using un-supervised pre-training with rectifier activations, contrary to what is experienced using tanh or soft-plus.Purely supervised rectifier networks remain competitive on all4datasets,even against the pretrained tanh or softplus models.3We also tested a rescaled version of the LIF and max(tanh(x),0)as activation functions.We obtained worse generalization performance than those of Table1, and chose not to report them.Xavier Glorot,Antoine Bordes,Yoshua Bengio•Rectifier networks are truly deep sparse networks.There is an average exact sparsity(fraction of ze-ros)of the hidden layers of83.4%on MNIST,72.0%on CIFAR10,68.0%on NISTP and73.8%on NORB.Figure3provides a better understand-ing of the influence of sparsity.It displays the MNIST test error of deep rectifier networks(with-out pre-training)according to different average sparsity obtained by varying the L1penalty on the works appear to be quite ro-bust to it as models with70%to almost85%of true zeros can achieve similar performances. With labeled data,deep rectifier networks appear to be attractive models.They are biologically credible, and,compared to their standard counterparts,do not seem to depend as much on unsupervised pre-training, while ultimately yielding sparse representations.This last conclusion is slightly different from those re-ported in(Nair and Hinton,2010)in which is demon-strated that unsupervised pre-training with Restricted Boltzmann Machines and using rectifier units is ben-eficial.In particular,the paper reports that pre-trained rectified Deep Belief Networks can achieve a test error on NORB below16%.However,we be-lieve that our results are compatible with those:we extend the experimental framework to a different kind of models(stacked denoising auto-encoders)and dif-ferent datasets(on which conclusions seem to be differ-ent).Furthermore,note that our rectified model with-out pre-training on NORB is very competitive(16.4% error)and outperforms the17.6%error of the non-pretrained model from Nair and Hinton(2010),which is basically what wefind with the non-pretrained soft-plus units(17.68%error).Semi-supervised setting Figure4presents re-sults of semi-supervised experiments conducted on the NORB dataset.We vary the percentage of the orig-inal labeled training set which is used for the super-vised training phase of the rectifier and hyperbolic tan-gent networks and evaluate the effect of the unsuper-vised pre-training(using the whole training set,unla-beled).Confirming conclusions of Erhan et al.(2010), the network with hyperbolic tangent activations im-proves with unsupervised pre-training for any labeled set size(even when all the training set is labeled). However,the picture changes with rectifying activa-tions.In semi-supervised setups(with few labeled data),the pre-training is highly beneficial.But the more the labeled set grows,the closer the models with and without pre-training.Eventually,when all avail-able data is labeled,the two models achieve identical performance.Rectifier networks can maximally ex-ploit labeled and unlabeledinformation.Figure4:Effect of unsupervised pre-training.On NORB,we compare hyperbolic tangent and rectifier net-works,with or without unsupervised pre-training,andfine-tune only on subsets of increasing size of the training set.4.2Sentiment AnalysisNair and Hinton(2010)also demonstrated that recti-fier units were efficient for image-related tasks.They mentioned the intensity equivariance property(i.e. without bias parameters the network function is lin-early variant to intensity changes in the input)as ar-gument to explain this observation.This would sug-gest that rectifying activation is mostly useful to im-age data.In this section,we investigate on a different modality to cast a fresh light on rectifier units.A recent study(Zhou et al.,2010)shows that Deep Be-lief Networks with binary units are competitive with the state-of-the-art methods for sentiment analysis. This indicates that deep learning is appropriate to this text task which seems therefore ideal to observe the behavior of rectifier units on a different modality,and provide a data point towards the hypothesis that rec-tifier nets are particarly appropriate for sparse input vectors,such as found in NLP.Sentiment analysis is a text mining area which aims to determine the judg-ment of a writer with respect to a given topic(see (Pang and Lee,2008)for a review).The basic task consists in classifying the polarity of reviews either by predicting whether the expressed opinions are positive or negative,or by assigning them star ratings on either 3,4or5star scales.Following a task originally proposed by Snyder and Barzilay(2007),our data consists of restaurant reviews which have been extracted from the restaurant review site .We have access to10,000 labeled and300,000unlabeled training reviews,while the test set contains10,000examples.The goal is to predict the rating on a5star scale and performance is evaluated using Root Mean Squared Error(RMSE).4 4Even though our tasks are identical,our database is。
背侧前额叶在工作记忆中的功能
i c 王贵振 李建民在进行学习、记忆、思维及问题解决等高级认知活动时,人 们需要一个暂时的信息加工与存储机制,它能够保存被激活的信息表征,以备进一步加工之用,1974年 B a d d e l e y 等[1]称这种 机制为工作记忆(w o r k i n g m e m o r y ),代表执行工作记忆的典型 反应为延迟反应。
在功能上,工作记忆包括执行控制和贮存激 活表征两个方面。
表征是暂时或永久储存在神经元网络里的 符号码。
从猴演示的延迟反应任务中得到的两个关键发现说明了前额叶(P F C )在工作记忆中的重要作用,其中尤其是背侧 前额 叶(d o r s l a t e r a l p r e f r o n t a l c o r t e x ,D L P F C )作 用。
首 先, D L P F C 主要沟回的损伤引起了延迟依赖的损害[2!4]。
其次,在 延迟反应任务的记忆间期,在 D L P F C 经常可以观察到持久固 定水平的神经冲动发放[5!7]。
本文主要通过以下三个方面来论述 D L P F C 在工作记忆中的作用进展。
1 D L P F C 功能模型明确的证据说明,延迟期间的活化事实上和储存相关,而不是 保持其他的一些相关过程。
如果延迟期间的活化反映了储存的表征,那么一个可能期 望就是记忆全过程的活化应该一直持续到它能指示一个反应。
在一些猴的 P F C 单元记录[5!7,13,18]事件相关f M R I 研究中,通过延长工作记忆任务中的记忆间期,已经报道 D L P F C 活性确实扩展到整个延迟阶段[19!22]。
这些结果和假设 D L P F C 在足够长的时间 内 储 存 活 化 的 表 征 来 指 导 合 适 的 行 为 是 一 致 的。
当然也可以这样解释,就是持续活化反映了把注意投向存在别 处的相关表征的进程。
另一个可能期望就是增加记忆负荷,延 迟期间的活化也应该增加。
无意识愤怒面孔加工的脑区激活似然估计(ALE)元分析
)
*+
)
!"# $%$"&’ (!(
"
$$ $%$!$"&"!
! "$ $ !& $%$(&’’(
左侧杏仁核对正性情绪起作用, 右侧杏仁核对负性 情绪起作用[24,25]。 两种假说从本研究看来均支持,首 先,右侧杏仁核的确在本研究结果中激活更明显,支
。。。。。。。。。。
。。。。。
。。。。。
。。。。。 。。。。。。
。 。 。 。。。。。 。。。 。。。 。 。 。 。
。。 。 。。 。 。Fra bibliotek 。 。 。 。
。。。。。
。。。。
。 。。。。。 。。。。。 。 。。。 。 。 。 。。
。。 。 。
。 。 。。 。 。 。 。
。。。。。
。。。。
。 。。。。。 。。。。。 。 。。。 。 。 。 。 。
·572·
Chinese Journal of Clinical Psychology Vol.19 No.5 2011
无意识愤怒面孔加工的脑区激活似然估计(ALE)元分析
姚树桥, 张江华, 石湖清 (中南大学湘雅二医院医学心理研究所,湖南 长沙 410011)
二甲双胍恩格列净片治疗2_型糖尿病合并HFpEF_的临床效果
- 6 -[7] NEERAJA A,PRASAD A H,OOMMEN M,et al.Evaluation and correlation of nociceptive response index and spectral entropy indices as monitors of nociception in anesthetized patients[J].Journal of Neurosciences in Rural Practice,2023,14(3):440-446.[8]赵曦,金毅,于砾淳,等.熵指数监测在神经内镜垂体瘤手术全身麻醉深度监测中的应用[J].现代肿瘤医学,2021,29(24):4388-4391.[9]徐海峰,王开俊,黄龙,等.熵指数在腹腔镜手术全身麻醉深度监测中的应用[J].中国现代药物应用,2021,15(10):78-80.[10]朱玉梅.熵指数监测在老年全身麻醉下行腹部手术患者中的应用[J].中国当代医药,2020,27(13):121-124.[11]陈玲坤,苏春燕,陈海林.熵指数在老年全身麻醉患者中的应用效果[J].吉林医学,2016,37(3):542-544.[12]张萱,高静.麻醉深度监测指标在围术期麻醉中的应用研究[J].医学信息,2022,35(14):168-171.[13]肯内特(Kenneth Zinzal).熵指数指导下不同剂量的右美托咪定对全凭静脉麻醉苏醒期的影响[D].通辽:内蒙古民族大学,2022.[14]李威,段春艳,高航.熵指数用于老年患者麻醉的临床效果观察[J].中国医药指南,2019,17(10):90-91.[15]邓大立,孙丽杰,崔艳,等.熵指数监测下不同全身麻醉方法用于腹腔镜胃癌根治术患者的效果比较[J/OL].临床医药文献电子杂志,2020,7(39):25.https:///kcms2/article/abstract?v=w1Je9LIFm5DsHfE0QofnujxMlnEUg9GzfCAm qGh8S-SdeUHwrMkLX-uswvmzWXSupu0_0dYVL-0zsDWyzZG FRhe1CaPzJZhoPOxAb4XO3lq2BudKvvqpqopZWPtZ64HmH7o DAPP9KYeIb8XjRmJt0Q==&uniplatform=NZKPT&language=C HS.[16]袁亮婧,金梅,张晓光.“三叶草”法超声引导下腰丛阻滞应用于髋关节镜术后镇痛的前瞻性随机对照研究[J].中国微创外科杂志,2019,19(11):964-968.[17]杨益锋,范智东.麻醉深度监测指标在围手术期麻醉中的应用研究进展[J].现代医药卫生,2021,37(12):2026-2029.[18]许克路,沈勤,柏耀林,等.腰丛阻滞复合不同麻醉深度喉罩全麻在老年THA 围手术期的应用[J].蚌埠医学院学报,2021,46(8):1017-1022.[19]谢秀秀,黄常君,赵建勇,等.镇痛指数与手术容积指数指导全麻镇痛药物使用的价值研究[J].浙江医学,2023,45(19):2080-2083,2088.[20]谢鹏程,李占芳,赵晓红,等.手术体积描计指数指导髋关节置换术中镇痛药使用的效果研究[J].临床麻醉学杂志,2019,35(5):451-453.(收稿日期:2024-02-02) (本文编辑:何玉勤)①承德医学院 河北 承德 067000②保定市第一中心医院全科医疗科 河北 保定 071000通信作者:杨晰晰二甲双胍恩格列净片治疗2型糖尿病合并HFpEF的临床效果杨玥①② 杨晰晰②【摘要】 目的:探讨二甲双胍恩格列净片治疗2型糖尿病(T2DM)合并射血分数保留性心力衰竭(HFpEF)患者的临床效果和安全性。
A_booming_field_of_large_animal_model_research
A booming field of large animal model researchAnimal models are integral to the study of fundamental biological processes and the etiology of human diseases.Small animal models, especially those involving mice, have yielded abundant and significant insights, greatly enhancing our understanding of biological phenomena and disease mechanisms. The preference for small animal models is primarily due to their ease of use and the availability of well-established genetic manipulation tools. The extensive use of genetically modified small animals has led to remarkable advancements in biomedical research. Nevertheless, it is important to recognize the substantial disparities in genomes,anatomy, and physiology between humans and small animals.As a result, there is an increasing awareness of the need to utilize large animal models that more closely resemble humans. Such models are essential for enhancing our understanding of crucial biological issues as well as the pathogenesis of diseases.This special column about “Large Animal Model Research ”,together with a previous review in this journal (Rahman et al.,2023), focusing specifically on large animal models and highlighting the use of various large animal models, including rabbits, pigs, and non-human primates (NHPs), in addressing key biological and medical issues. Incorporating larger animals into research not only substantially enhances the variety of research tools and methodologies available, but also enables the exploration of long-standing issues that have proven challenging using small animals, given their species-specific differences.Among the abovementioned large animals, rabbits are comparatively less commonly employed as a model system for research. However, the comprehensive review by Han et al. (2024b) offers valuable insights into the utilization of rabbit models in the field of biomedical research. Rabbits represent a cost-effective advantage over larger animals,owing to their ease of handling and rapid reproductive capacity, while also exhibiting larger body sizes and longer lifespans than rodents. These unique attributes, coupled with advancements in genetic modification techniques, such as CRISPR/Cas9 gene targeting, have led to the establishment of various rabbit models for biomedical research. Given that their lipid metabolic profiles and immune responses are more similar to humans than to rodents, rabbits are considered suitable for modeling cardiovascular and immune-related diseases. Furthermore, with their extensive involvement in commercial antibody production, genetically modified rabbits are an asset in the development of antibody-based drugs andimmunotherapeutic agents. Although the generation of rabbit models dedicated to the exploration of neurological disorders has been relatively limited, a recent study using base editing technology to modify the amyotrophic lateral sclerosis (ALS)gene SOD1 successfully produced a rabbit model exhibiting ALS phenotypes (Zhang et al., 2023). Thus, this study underscores the potential of rabbits in investigating neurological disorders.Regarding neurological disorder research, NHPs play an indispensable role due to their high similarity to humans in terms of brain structure, function, and aging processes. In their comprehensive review, Pan et al. (2024) discuss several NHP models of neurodegenerative diseases, including Alzheimer’s disease (AD), Parkinson’s disease (PD), ALS, and Huntington’s disease (HD). The analysis of these disease models underscores the significance of employing NHPs for the investigation of the fundamental mechanisms underpinning neurodegenerative diseases.Despite their extensive use over the past few decades,genetically modified mouse models cannot fully replicate the pathological features observed in human patients. A notable discrepancy is the absence of pronounced and selective neuronal loss in most genetically modified mouse models. For instance, mouse models carrying genetic mutations associated with PD do not exhibit degeneration of dopaminergic neurons, and most mouse models of ALS lack significant cytoplasmic accumulation of TDP-43, both of which are key pathological manifestations observed in human PD and ALS, respectively. In contrast, monkeys expressing PD-associated proteins (Li et al., 2021a; Yang et al., 2019) or mutant TDP-43 (Yin et al., 2019) successfully recapitulate neurodegeneration and cytoplasmic TDP-43 accumulation,respectively.Neurodegenerative features are not limited to primates; they also occur in other types of large animals. Pigs have been utilized in the exploration of neurological disorders due to their genetic, anatomical, and physiological similarities to humans.In the review by Han et al. (2024a), pigs, NHPs, and sheep were compared for their utility in studying HD, a monogenetic disorder with a genetic mutation that can be replicated across different species. Similar to mouse models of AD and PD,mouse models carrying the HD gene also fail to exhibit the robust and overt neurodegeneration observed in afflicted patient brains. Several transgenic large animal models have been established in NHPs, pigs, and sheep. Nevertheless,these models present a spectrum of phenotypic variations,and the progression and severity of disease differ significantly.Given that animal behaviors and phenotypes are strongly influenced by transgene copy number and expression levels,ideal animal models are those that can faithfully recapitulate human genetic mutations and express mutant genes at the endogenous level.Pigs possess distinct genetic modification advantages compared to NHPs and sheep. Notably, somatic nuclearThis is an open-access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (/licenses/by-nc/4.0/), which permits unrestricted non-commercial use, distribution, and reproduction in any medium,provided the original work is properly cited.Copyright ©2024 Editorial Office of Zoological Research, Kunming Institute of Zoology, Chinese Academy of SciencesLi & Lai. Zool. Res. 2024, 45(2): 311−313https:///10.24272/j.issn.2095-8137.2024.018transfer, which allows for the knock-in (KI) of mutant genes, has been successfully applied in pigs to generate a HD gene KI model that recapitulates selective neuronal loss and motor function deficiency, two crucial pathological features of HD. This HD KI pig model has been effectively employed in gene therapy research to reduce HD pathology and symptoms (Yan et al., 2023), providing further evidence for the importance of large animal models in the investigation of neurodegenerative diseases.Pigs are also extensively used in other biomedical research fields, including circulatory system diseases, organ transplantation, diabetes, skin diseases, and tumors. Porcine models also show promise for studying inherited hearing loss. Notably, pigs have a hearing range comparable to that of humans and naturally undergo age-related hearing loss, a phenomenon largely attributed to the remarkable similarity in both the structure and function of their ears to those of humans. Genetic modification techniques, such as CRISPR/Cas9 gene editing and somatic nuclear transfer, have enabled the creation of single genetic mutations in pig models, facilitating the study of disease pathogenesis in monogenic hearing loss. For example, Wang et al. (2024) discussed the use of pig models in studying hearing loss-related diseases, with notable implications for other monogenic diseases.The four research articles published in this special issue also highlight the significance of large animal models in research. Li et al. (2024a) reported on mitochondrial replacement in cynomolgus monkeys (Macaca fascicularis) through female pronuclear transfer. Their findings suggest that pronuclear transfer holds great potential in reducing the risk of inherited mitochondrial DNA (mtDNA) diseases, thereby establishing the applicability of non-human models in investigating diseases related to mtDNA defects. Li et al. (2024b) developed cynomolgus monkey organoids to study neural tube defects (NTDs), which can also arise from loss-of-function mutations in the SHROOM3 gene. Their findings indicate that in vitro models using NHP-derived organoids can be employed to investigate the pathogenesis of significant human diseases.The rationale for using NHPs in research is also based on species-dependent gene expression patterns. Chen et al. (2024) presented intriguing findings regarding the expression of PINK1, which is associated with PD, in mice, pigs, and monkeys. Using multiple antibodies and PINK1 knockout animal models, they discovered that the PINK1 protein, rather than PINK1 mRNA, is detectably and exclusively expressed in the primate brain. These findings explain why PINK1 knockout leads to neurodegeneration in monkeys, but not in mice and pigs, suggesting that a possible higher abundance of endogenous PINK1 in primate brains contributes to neuronal survival. Their results also imply that species-dependent transcriptional and translational regulations contribute to species-specific pathology. Therefore, utilizing animal models that more closely resemble humans provides better prospects for uncovering the molecular mechanisms underlying primate-specific gene expression. Reinforcing this idea, Mao et al. (2024) conducted a comprehensive analysis of the transcriptomes within macaque species and between macaques and humans, highlighting the high conservation of tissue-specific genes and the value of macaques as biological models for investigating human diseases. Additionally, the identification of a cynomolgus monkey with naturally occurring PD further supports the suitability of NHPs as ideal models for PD research (Li et al., 2021b).Despite the significant contributions and advancements in large animal models for biomedical research, notable limitations and challenges exist. The utilization of large animals necessitates stricter regulation and control due to ethical concerns arising from their high cognitive capacities and close resemblance to humans. Additionally, the substantial costs and extended breeding periods associated with maintaining large animals pose considerable challenges in expanding the scope of studies using these models. Undoubtedly, small animal models, particularly rodents, are likely to remain fundamental in biological and biomedical research, continuing to play a critical role in enhancing our understanding of disease pathogenesis and in developing therapeutic strategies for human diseases. However, in scenarios where small animal models have failed to replicate important pathological events observed in humans, it is worth considering large animal alternatives to unravel disease mechanisms that may not be evident in small animals. Furthermore, large animal models can facilitate the validation of specific therapeutic targets, potentially reducing the failure rates of clinical trials for therapeutics initially developed using small animal models. The integration of large animal models into research endeavors promises to deepen our understanding of complex biological processes and improve the translation of preclinical findings into clinical practice.Xiao-Jiang Li1,2,*, Liangxue Lai3,*1 Guangdong Key Laboratory of Non-human Primate Research,Key Laboratory of CNS Regeneration (Ministry of Education), GHM Institute of CNS Regeneration, Jinan University,Guangzhou, Guangdong 510632, China2 State Key Laboratory of Bioactive Molecules and Druggability Assessment, Jinan University, Guangzhou, Guangdong 510632,China3 CAS Key Laboratory of Regenerative Biology, GuangdongProvincial Key Laboratory of Stem Cell and RegenerativeMedicine, Guangzhou Institutes of Biomedicine and Health, Chinese Academy of Sciences, Guangzhou, Guangdong 510530,China*Corresponding authors, E-mail: **************.cn;********************.cnREFERENCESChen XS, Han R, Liu YT, et al. 2024. Comparative analysis of primate and pig cells reveals primate-specific PINK1 expression and phosphorylation. Zoological Research, 45(2): 242−252.Han B, Liang W, Li XJ, et al. 2024a. Large animal models for Huntington’s disease research. Zoological Research, 45(2): 275−283.Han Y, Zhou J, Zhang R, et al. 2024b. Genome-edited rabbits: Unleashing the potential of a promising experimental animal model across diverse diseases. Zoological Research, 45(2): 253−262.Li CY, Liu XC, Li YZ, et al. 2024a. Generation of mitochondrial replacement monkeys by female pronucleus transfer. Zoological Research, 45(2): 292−298.Li H, Wu SH, Ma X, et al. 2021a. Co-editing PINK1 and DJ-1 genes via adeno-associated virus-delivered CRISPR/Cas9 system in adult monkey brain elicits classical parkinsonian phenotype. Neuroscience Bulletin, 37(9): 1271−1288.Li H, Yao YG, Hu XT. 2021b. Biological implications and limitations of a312 cynomolgus monkey with naturally occurring Parkinson's disease. Zoological Research, 42(2): 138−140.Li P, Zhang T, Wu R, et al. 2024b. Loss of SHROOM3 affects neuroepithelial cell shape through regulating cytoskeleton proteins in cynomolgus monkey organoids. Zoological Research, 45(2): 233−241.Mao YX, Li Y, Yang Z, et al. 2024. Comparative transcriptome analysis between rhesus macaques (Macaca mulatta) and crab-eating macaques (Macaca fascicularis). Zoological Research, 45(2): 299−310.Rahman A, Li Y, Chan TK, et al. 2023. Large animal models of cardiac ischemia-reperfusion injury: Where are we now?. Zoological Research, 44(3): 591−603.Pan MT, Zhang H, Li XJ, et al. 2024. Genetically modified non-human primate models for research on neurodegenerative diseases. Zoological Research, 45(2): 263−274.Wang X, Liu TX, Zhang Y, et al. 2024. Genetically modified pigs: Emerging animal models for hereditary hearing loss. Zoological Research, 45(2): 284−291.Yan S, Zheng X, Lin YQ, et al. 2023. Cas9-mediated replacement of expanded CAG repeats in a pig model of Huntington's disease. Nature Biomedical Engineering, 7(5): 629−646.Yang WL, Liu YB, Tu ZC, et al. 2019. CRISPR/Cas9-mediated PINK1 deletion leads to neurodegeneration in rhesus monkeys. Cell Research, 29(4): 334−336.Yin P, Guo XY, Yang WL, et al. 2019. Caspase-4 mediates cytoplasmic accumulation of TDP-43 in the primate brains. Acta Neuropathologica, 137(6): 919−937.Zhang ZT, Wu XY, Yang J, et al. 2023. Highly efficient base editing in rabbit by using near-PAMless engineered CRISPR/Cas9 variants. Science China Life Sciences, 66(3): 635−638.Zoological Research 45(2): 311−313, 2024 313。
2022年考研英语一答案解析(完整版)
2022英语一考研真题(含答案)Section I Use of EnglishDirections: Read the following text. Choose the best word(s) for each numbered blank and mark [A], [B]. [C].or [D] on the ANSWER SHEET. (10 points) The idea that plants have some degree of consciousness first took root in the early 2000s: the term "plant neurobiology was (1) around the notion that some aspects of plant behavior could be (2) to intelligence in animals. (3) plants lack brains, the firing of electrical signals in their stems and leaves nonetheless triggered responses that (4) consciousness, researchers previously reported.But such an idea is untrue, according to a new opinion article. Plant biology is complex and fascinating, but it (5) so greatly from that of animals that so-called (6) of plants intelligence is inconclusive, the authors wrote.Beginning in 2006, some scientists have (7) that plants possess neuron-like cells that interact with hormones and neurotransmitters, (8) "a plant nervous system, (9) to that in animals."said lead study author Lincoln Taiz, "They (10) claimed that plants have "brain-like command centers"at their root tips."This (11) makes sense if you simplify the workings of a complex brain. (12) it to an array of electrical pulses; cells in plants also communicate through electrical signals. (13) , the signaling in a plant is only (14) similar to the firing in a complex animal brain, which is more than "a mass of cells that communicate by electricity."Taiz said"For consciousness to evolve, a brain with a threshold (15) of complexity and capacity is required,"he (16) "Since plants don't have nervous systems, the (17) that they have consciousness are effectively zero."And what's so great about consciousness, anyway? Plants can't run away from (18) . so investing energy in a body system which (19) a threat and can feel pain would be a very (20) evolutionary strategy, according to the article.1.A.coined B.discovered C.collected D.issued2.A.attributed B.directed pared D.confined3.A.unless B.when C.once D.though4.A.cope with B.consisted of C.hinted at D.extended in5.A.suffers B.benefits C.develops D.differs6.A.acceptance B.evidence C.cultivation D.creation7.A.doubted B.denied C.argued D.requested8.A.adapting B.forming C.repairing D.testing9.A.analogous B.essential C.suitable D.sensitive10.A.just B.ever C. still D.even11.A.restriction B.experiment C.perspective D.demand12.A.attaching B.reducing C.returning D.exposing13.A.However B.Moreover C.Therefore D.Otherwise14.A.temporarily B.literally C.superficially D.imaginarily15.A.list B.level bel D.local16.A.recalled B.agreed C.questioned D.added17.A.chances B.risks C.excuses D.assumptions18.A. danger B.failure C.warning D.control19.A.represents B.includes C.reveals D.recognizes20.A.humble B.poor C.practical D.easySection II Reading ComprehensionPart A Directions: Read the following four texts. Answer the questions below each text by choosing [A], [B]. [C], or [D].Mark your answers on the ANSWER SHEET. (40 points)Text 1People often complain that plastics are too durable. Water bottles. shopping bags, and other trash litter the planet, from Mount Everest to the Mariana Trench, because plastics are everywhere and don't break down easily. But some plastic materials change over time. They crack and frizzle.They "weep"out additives. They melt into sludge. All of which creates huge headaches for institutions, such as museums, trying to preserve culturally important objects. The variety of plastic objects at risk is dizzying: early radios, avant-garde sculptures, celluloid animation stills from Disney films, the first artificial heart.Certain artifacts are especially vulnerable because some pioneers in plastic art didn't always know how to mix ingredients properly, says Thea van Oosten, a polymer chemist who, until retiring a few years ago, worked for decades at the Cultural Heritage Agency of the Netherlands."It' s like baking a cake: If you don't have exact amounts, it goes wrong,"she says. "The object you make is already a time bomb"And sometimes, it's not the artist's fault. In the 1960s. the Italian artist Picro Gilardi began to create hundreds of bright, colorful foam pieces. Those pieces included small beds of roses and other items as well as a few dozen "nature carpets"-large rectangles decorated with foam pumpkins, cabbages, and watermelons. He wanted viewers to walk around on the carpets -which meant they had to be durable.Unfortunately, the polyurethane foam he used is inherently unstable. It's especially vulnerable to light damage, and by the mid-1990s, Gilardi's pumpkins, roses, and other figures were splitting and crumbling, Museums locked some of them away in the dark. So van Oosten and her colleagues worked to preserve Gilardi's sculptures. They infused some with stabilizing and consolidating chemicals. Van Oosten calls those chemicals "sunscreens"because their goal was to prevent further light damage and rebuild worn polymer fibers. She is proud that several sculptures have even gone on display again, albeit sometimes beneath protective cases.Despite success stories like van Oosten' s, preservation of plastics will likely get harder. Old objects continue to deteriorate. Worse, biodegradable plastics designed to disintegrate, are increasingly common. And more is at stake here than individual objects. Joana Lia Ferreira. an assistant professor of conservation and restoration at the NOV A School of Science and Technology notes that archaeologists first defined the great material ages of human history -Stone Age, Iron Age, and so on -after examining artifacts in museums. We now live in an age of plastic, she says,"and what we decide to collect today, what we decide to preserve... will have a strong impact on how in the future we' ll be seen.21. According to Paragraph 1, museums are faced with difficulties in .[A] maintaining their plastic items[B] obtaining durable plastic artifacts[C] handling outdated plastic exhibits[D] classifying their plastic collections22. Van Oosten believes that certain plastic objects are .[A] immune to decay[B] improperly shaped[C] inherently flawed[D] complex in structure23. Museums stopped exhibiting some of Gilardi's artworks to .[A] keep them from hurting visitors[B] duplicate them for future display[C] have their ingredients analyzed[D] prevent them from further damage24. The author thinks that preservation of plastics is .[A] costly[B] unworthy[C] unpopular[D] challenging25. In Ferreira' s opinion, preservation of plastic artifacts .[A] will inspire future scientific research[B] has profound historical significance[C] will help us separate the material ages[D] has an impact on today's cultural lifeText 2As the latest crop of students pen their undergraduate applications and weigh up their options, it may be worth considering just how the point, purpose and value of a degree has changed and what Gen Z need to consider as they start the third stage of their educational journey.Millennials were told that if you did well in school. got a decent degree, you would be set up for life. But that promise has been found wanting. As degrees became universal, they became devalued. Education was no longer a secure route of social mobility. Today, 28 per cent of graduates in the UK are in non-graduate roles: a percentage which is double the average amongst the OECD.This is not to say that there is no point in getting a degree, but, rather stress that a degree is not for everyone, that the switch from classroom to lecture hall is not an inevitable one and that other options are available.Thankfully, there are signs that this is already happening, with Gen Z seeking to learn from their millennial predecessors, even if parents and teachers tend to be still set in the degree mindset.Employers have long seen the advantages of hiring school leaverswho often prove themselves to be more committed and loyal employees than graduates. Many too are seeing the advantages of scrapping a degree requirement for certain roles.For those for whom a degree is the desired route, consider that this may well be the first of many. In this age of generalists, it pays to have specific knowledge or skills. Postgraduates now earn 40 per cent more than graduates. When more and more of us have a degree, it makes sense to have two.It is unlikely that Gen Z will be done with education at 18 or 21; they will need to be constantly up-skilling throughout their career to stay agile, relevant and employable. It has been estimated that this generation due to the pressures of technology, the wish for personal fulfilment and desire for diversity will work for 17 different employers over the course of their working life and have five different careers. Education, and not just knowledge gained on campus, will be a core part of Generation Z's career trajectory.Older generations often talk about their degree in the present and personal tense: I am a geographer' or T am a classist. Their sons or daughters would never say such a thing: it's as if they already know that their degree won't define them in the same way.26. The author suggests that Generation Z should .[A] be careful in choosing a college[B] be diligent at each educational stage[C] reassess the necessity of college education[D] postpone their undergraduate application27. The percentage of UK graduates in non-graduate roles reflect .[A] Millennial's opinions about work[B] the shrinking value of a degree[C] public discontent with education[D] the desired route of social mobility28. The author considers it a good sign that .[A] Generation Z are seeking to earn a decent degree[B] school leavers are willing to be skilled workers[C] employers are taking a realistic attitude to degree[D] parents are changing their minds about education29. It is advised in Paragraph 5 that those with one degree should .[A] make an early decision on their career[B] attend on the job training programs[C] team up with high-paid postgraduates[D] further their studies in a specific field30. What can be concluded about Generation Z from the last two paragraphs?[A] Lifelong leaning will define them.[B]They will make qualified educators.[C] Depress will no longer appeal them.[D] They will have a limited choice of jobs.Text 3Enlightening, challenging, stimulating, fum. These were some of the words that Nature readers used to describe their experience of art-science collaborations in a series of articles on partnerships between artists and researchers. Nearly 40% of the roughly 350 people who responded to an accompanying poll said, they had collaborated with artists: and almost all said they would consider doing so in future.Such an encouraging results is not surprising. Scientists are increasingly seeking out visual artists to help them communicate their work to new audiences. "Artists help scientists reach a broader audience and make emotional connections that enhance learning."One respondent said.One example of how artists and scientists have together rocked the scenes came last month when the Sydney Symphony Orchestra performed a reworked version of Antonio Vivaldi's The Four Seasons. They reimagined the 300-year-old score by injecting the latest climate prediction data for each season-provided by Monash University's Climate Change Communication Research Hub. The performance was a creative call to action ahead of November' s United Nations Climate Change Conference in Glasgow, UK.But a genuine partnership must be a two-way street. Fewer artist than scientists responded to the Nature poll, however, several respondents noted that artists do not simply assist scientists with their communication requirements. Nor should their work be considered only as an object of study. The alliances are most valuable when scientists and artists have a shared stake in a project, are able to jointly design it and can critique each other' s work. Such an approach can both prompt new research as well as result in powerful art. More than half a century ago, the Massachusetts Institute of Technology opened its Center for Advanced Visual Studies (CA VS) to explore the role of technology in culture. The founders deliberately focused their projects around light-hance the"visual studies"in the name. Light was a something that both artists and scientists had an interest in, and therefore could form the basis of collaboration. As science and technology progressed, and divided into more sub-disciplines, the centre was simultaneously looking to a time when leading researchers could also be artists. writers and poets, and vice versa.Nature' s poll findings suggest that this trend is as strong as ever, but, to make a collaboration work, both sides need to invest time, and embrace surprise and challenge. The reach of art-science tie-ups needs to go beyond the necessary purpose of research communication, and participants.Artists and scientists alike are immersed in discovery and invention, and challenge and critique are core to both, too.31. According to paragraph 1, art-science collaborations have .[A] caught the attention of critics[B] received favorable responses[C] promoted academic publishing[D] sparked heated public disputes32. The reworked version of The Four Seasons is mentioned to show that .[A] art can offer audiences easy access to science[B] science can help with the expression of emotions[C] public participation in science has a promising future[D] art is effective in facilitating scientific innovations33. Some artists seem to worry about in the art-science partnership .[A] their role may be underestimated[B] their reputation may be impaired[C] their creativity may be inhibited[D] their work may be misguided34. What does the author say about CA VS?[A] It was headed alternately by artists and scientists.[B] It exemplified valuable art-science alliances.[C] Its projects aimed at advancing visual studies.[D] Its founders sought to raise the status of artists.35. In the last paragraph, the author holds that art-science collaborations .[A] are likely to go beyond public expectations[B] will intensify interdisciplinary competition[C] should do more than communicating science.[D] are becoming more popular than beforeText 4The personal grievance provisions of New Zealand's Employment Relations Act 2000 (ERA)prevent an employer from firing an employee without good cause. Instead. dismissals must be justified. Employers must both show cause and act in a procedurally fair way.Personal grievance procedures were designed to guard the jobs of ordinary workers from "unjustified dismissals". The premise was that the common law of contract lacked sufficient safeguards for workers against arbitrary conduct by management. Long gone are the days when a boss could simply give an employee contractual notice.But these provisions create difficulties for businesses when applied to highly paid managers and executives. As countless boards and business owners will attest, constraining firms from firing poorly performing. high-earning managers is a handbrake on boosting productivity and overall performance. The difference between C-grade and A-grade managers may very well be the difference between business success or failure. Between preserving the jobs of ordinary workers or losing them. Yet mediocrity is no longer enough to justify a dismissal.Consequently—and paradoxically—laws introduced to protect the jobs of ordinary workers may be placing those jobs at risk If not placing jobs at risk, to the extent employment protection laws constrain business owners from dismissing under-performing managers, those laws act as a constraint on firm productivity and therefore on workers' wages. Indeed, in "An International Perspective on New Zealand's Productivity Paradox"(2014). the Productivity Commission singled out the low quality of managerial capabilities as a cause of the country' s poor productivity growth record.Nor are highly paid managers themselves immune from the harm caused by the ERA's unjustified dismissal procedures. Because employment protection laws make it costlier to fire an employee, employers are more cautious about hiring new staff. This makes it harder for the marginal manager to gain employment. And firms pay staff less because firms carry the burden of the employment arrangement going wrong.Society also suffers from excessive employment protections. Stringent job dismissal regulations adversely affect productivity growth and hamper both prosperity and overall well-being.Across the Tasman Sea, Australia deals with the unjustified dismissal paradox by excluding employees earning above a specified "high-income threshold"from the protection of its unfair dismissal laws. In New Zealand, a 2016 private members' Bill tried to permit firms and high-income employees to contract out of the unjustified dismissal regime. However, the mechanisms proposed were unwieldy and the Bill was voted down following the change in government later that year.36. The personal grievance provisions of the ERA are intended to .[A] punish dubious corporate practices[B] improve traditional hiring procedures[C] exempt employers from certain duties[D] protect the rights of ordinary workers37. It can be learned from Paragraph 3 that the provisions may .[A] hinder business development[B] undermine managers' authority[C] affect the public image of the firms[D] worsen labor-management relations38. Which of the following measures would be the Productivity Commission support?[A] Imposing reasonable wage restraints.[B] Enforcing employment protection laws[C] Limiting the powers of business owners.[D] Dismissing poorly performing managers.39. What might be an effect of ERA's unjustified dismissal procedures?[A] Highly paid managers lose their jobs.[B] Employees suffer from salary cuts.[C] Society sees a rise in overall well-being.[D] Employers need to hire new staff.40. It can be inferred that the "high-income threshold"in Australia .[A] has secured managers' earnings[B] has produced undesired results[C] is beneficial to business owners[D] is difficult to put into practicePart BDirections:In the following text:some sentences have been removed.For Questions 41-45:choose the most suitable one from the fist A-G to fit into each of the numbered blanks.There are two extra choices:which do not fit in any of the gaps.Mark your answers on ANSWER SHEET.:10 points:——7选5(41)Teri Byr dI was a zoo and wildlife park employee for years. Both the wildlife park and zoo claimed to be operating for the benefit of the animals and for conservation purposes. This claim was false.Neither one of them actually participated in any contributions whose bottom line is much more important than the condition of the animals.Animals despise being captives in zoos. No matter how you "enhance"enclosures, they do not allow for freedom, a natural diet or adequate time for transparency with these institutions. and it's past time to eliminate zoos from our culture.(42)Karen R. SimeAs a zoology professor, I agree with Emma Marris that zoo displays can be sad and cruel. But she underestimates the educational value of zoos.The zoology program at my university attracts students for whom zoo visits were the crucial formative experience that led them to major in biological sciences. These are mostly students who had no opportunity as children to travel to wilderness areas, wildlife refuges or national parks. Although good TV shows can help stir children's interest in conservation, they cannot replace the excitement of a zoo visit as an intense, immersive and interactive experience. Surely there must be some middle ground that balances zoos' treatment of animals with their educational potential.(43)Greg NewberryEmma Marris's article is an insult and a disservice to the thousands of passionate who work tirelessly to improve the lives of animals and protect our planet. She uses outdated research and decades-old examples to undermine the noble mission of organization committed to connecting children to a world beyond their own.Zoos are at the forefront of conservation and constantly evolving to improve how they care for animals and protect each species in its natural habitat. Are there tragedies? Of course. But they are the exception not the norm that Ms. Marris implies. A distressed animal in a zoo will get as good or better treatment than most of us at our local hospital(44) Dean GalleaAs a fellow environmentalist animal-protection advocate and longtime vegetarian.I could properly be in the same camp as Emma Marris on the issue of zoos. But I believe that well-run zoos and the heroic animals that suffer their captivity so serve a higher purpose. Were it not for opportunities to observe these beautiful wild creatures close to home many more people would be driven by their fascination to travel to wild areas to seek out disturb and even hunt them down.Zoos are in that sense similar to natural history and archeology museums serving to satisfy our need for contact with these living creatures while leaving the vast majority undisturbed in their natural environments.(44)John FraserEmma Marris selectively describes and misrepresents the findings of our research. Our studies focused on the impact of zoo experiences on how people think about themselves and nature and the data points extracted from our studies.Zoos are tools for thinking. Our research provides strong support for the value of zoos in connecting people with animals and with nature. Zoos provide a critical voice for conservation and environmental protection. They afford an opportunity for people from all backgrounds to encounter a range of animals from drone bees to springbok or salmon to better understand the natural world we live in.A.Zoos which spare no effort to take care of animals should not be subjected to unfair criticismB.To pressure zoos to spend less on their animals would lead to inhumane outcomes for the precious creatures in their care.C.While animals in captivity deserve sympathy, zoos play a significant role in starting young people down the path of related sciences.D.Zoos save people trips to wilderness areas and thus contribute to wildlife conservation.E.For wild animals that cannot be returned to their natural habitats, zoos offer the best alternative.F.Zoos should have been closed down as they prioritize money making over animals' wellbeing.G.Marris distorts our findings which actually prove that zoos serve as an indispensable link between man and nature.答案:41.F 42.C 43.A 44.D 45.GPart CDirections: Read the following text carefully and then translate the underlined segments into Chinese. Your translation should be written neatly on the ANSWER SHEET. (10 points)"Between 1807 and 1814 the Iberian Peninsula (comprising Spain and Portugal) was the scene of a titanic and merciless struggle. It took place on many different planes: between Napoleon's French army and the angry inhabitants: between the British, ever keen to exacerbate the emperor's difficulties, and the marshals sent from Paris to try to keep them in check: between new forces of science and meritocracy and old ones of conservatism and birth.(46) It was also, and this is unknown even to many people well read about the period. a battle between those who made codes and those who broke them.I first discovered the Napoleonic cryptographic battle a few years ago when I was reading Sir Charles Oman's epic History of the Peninsular War. In volume V he had attached an appendix. "The Scovell Ciphers."(47) It listed many documents in code that had been captured from the French army of Spain, and whose secrets had been revealed by the work of one George Scovell, an officer in British headquarters. Oman rated Scovell's significance highly, but at the same time, the general nature of his History meant that (48) he could not analyze carefully what this obscure officer may or may not have contributed to that great struggle between nations or indeed tell us anything much about the man himself. I was keen to read more, but was surprised to find that Oman's appendix, published in 1914. was the only considered thing that had been written about this secret ware.I became convinced that this story was every bit as exciting and significant as that of Enigma and the breaking of German codes in the Second World War. The question was, could it be told?Studying Scovell's papers at the Public Record Office (in Kew., west London) I found that he had left an extensive journal and copious notes a work in the peninsula. What was more, many original French dispatches had been preserved in this collection.I realized at once that this was priceless. (49) There may have been many spies and intelligence officers during the Napoleonic Wars, but it is usually extremely difficult to find the material they actually provided or worked on. Furthermore. Scovell's storyinvolved much more than just intelligence work. His status in Lord Wellington's headquarters and the recognition given to him for his work were all bound up with the class politics of the army at the time. His tale of self-improvement and hard work would make a fascinating biography in its own right, but represents something more than that.(50) Just as the code breaking has its wider relevance in the struggle for Spain, so his attempts to make his way up the promotion ladder speak volumes about British society.The story of Wellington himself also gripped me. Half a century ago his campaigns were considered a central part of the British historical mythology and spoon-fed to schoolboys.More recently this has not been the case, which is a great shame. A generation has grown up.46.甚至是许多熟悉这段时期的人都不知道,这也是一场密码制作者与破解者之间的斗争。
Circuit-specific_gene_therapy_is_knocking_on_the_d
Circuit-specific gene therapy is knocking on the door of Parkinson’s diseaseParkinson’s disease (PD) is a prevalent neurodegenerative disorder that affects millions of individuals worldwide.Symptoms of PD typically manifest as movement impairments,including bradykinesia, rigidity, tremors, and postural instability, as well as non-motor symptoms, such as cognitive decline, pain, and depression (Bloem et al., 2021). The fundamental neuropathological hallmarks of PD include the degeneration of dopaminergic neurons in the substantia nigra (SN) and the aggregation of α-synuclein in intracellular inclusions. At present, the primary interventions for PD treatment are levodopa, dopamine agonists, deep brain stimulation (DBS), and physical therapy. However, levodopa administration can lead to intractable side effects such as dyskinesia (Wang & Shih, 2023) and DBS poses risks of long-term psychiatric complications, including chronic depression and personality alterations, as well as infection related to the intracranial electrodes (Volkmann et al., 2010). Consequently,there exists an urgent need for the innovation of targeted and efficacious treatments for PD.Chen et al. (2023) recently documented the development and verification of a novel gene therapy for PD, which successfully reversed the core motor symptoms of PD in both mouse and monkey models (Figure 1). Their research specifically targeted D1 medium spiny neurons (D1-MSNs) in the striatum, which project to the globus pallidus internal segment (GPi) and substantia nigra pars reticulata (SNr), thus forming a direct pathway. D2 medium spiny neurons (D2-MSNs), another important neuronal type in the striatum,project indirectly to the SNr via the globus pallidus external segment (GPe)/subthalamic nucleus (STN), thus forming an indirect pathway. Activation of the direct pathway (comparable to an “accelerator”) facilitates movement, whereas activation of the indirect pathway (comparable to a “brake”) inhibits movement. In PD, degeneration of dopaminergic neurons in the substantia nigra pars compacta (SNc) weakens dopaminergic modulation of both D1-MSNs and D2-MSNs. As a result, the direct pathway is repressed while the indirect pathway is enhanced, causing an imbalance between the “accelerator” and “brake”, resulting in motor deficits.Therefore, the study aimed to up-regulate the activity of D1-MSNs associated with the direct pathway and thereby restore movement control equilibrium in PD (Chen et al., 2023).To achieve this objective, a method to precisely manipulatespecific cell types within D1-MSNs is required. Thus, the authors utilized cell type-specific expression by employing a tailored retrograde adeno-associated virus (AAV) with a unique promoter to drive elevated gene expression in D1-MSNs, the only major cell type in the striatum projecting to the SNr. Specifically, the team first engineered several AAV capsid mutants, with the AAV8 mutant, AAV8R, showing improved retrograde infection of D1-MSNs. Subsequent modifications led to the identification of AAV8R12 as the most efficient mutant for labeling D1-MSNs. Next, the authors analyzed a gene expression database and screened for promoters. Through extensive in vivo experimentation, the authors ascertained that the G88 promoter family drove the highest level of gene expression in MSNs. Consequently, by integrating the G88P3/G88P7 promoter with the AAV8R12vector, they obtained a robust manipulation tool capable of labeling D1-MSNs with strong specificity in primates.Next, to modulate the functions of the labeled D1-MSNs, the authors expressed Designer Receptors Exclusively Activated by Designer Drugs (DREADD) effector rM3Ds into D1-MSNs by unilaterally injecting the viral vector AAV8R12-G88P7-rM3Ds -2A -EYFP into the SNr of monkeys. Following the administration of clozapine N-oxide (CNO) and deschloroclozapine (DCZ), ligands that bind with rM3Ds to enable neuronal excitation (Roth, 2016), the monkeys showed marked increases in contraversive rotations, thus demonstrating the efficiency of the toolkit in direct pathway activation.Finally, the same approach was applied to MPP +-treated monkeys, which exhibited PD-like symptoms, including tremor,bradykinesia, and rigidity. Following a single surgical procedure to inject the virus into the SNr, the monkeys receiving regular CNO/DCZ treatments showed rapid recovery and restored motor ability, including increased spontaneous movement, reduced tremor, and improved motor skills.Continuous monitoring of these subjects confirmed the long-term safety of the treatment. Notably, when compared to the standard therapeutic dose of levodopa, the most widely used treatment for PD, this circuit-specific gene therapy not only showed a more rapid reversal of symptoms but also no inducement of observable side effects.Chen et al. (2023) not only established a proof-of-concept for circuit-specific therapeutics, but also effectively leveraged a non-human primate (NHP) model to fully showcase theReceived: 07 November 2023; Accepted: 15 November 2023; Online: 17November 2023Foundation items: This work was supported by the National Natural Science Foundation of China (32160204), Major Science and Technology Projects of Hainan Province (ZDKJ2021032), and STI 2030—Major Projects (2022ZD0208602)This is an open-access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (/licenses/by-nc/4.0/), which permits unrestricted non-commercial use, distribution, and reproduction in any medium,provided the original work is properly cited.Copyright ©2023 Editorial Office of Zoological Research, Kunming Institute of Zoology, Chinese Academy of SciencesWang et al. Zool. Res. 2023, 44(6): 1152−1153https:///10.24272/j.issn.2095-8137.2023.353effectiveness and safety of the therapy in PD monkeys. Given the natural occurrence of parkinsonian conditions in monkeys (Li et al., 2021a, 2021b), indicating shared mechanisms and pathologies of PD between NHPs and humans, this approach holds promise for successful clinical translation and application. In contrast, numerous proposed therapeutics have claimed efficacy in alleviating PD symptoms in mice. However, the significant physiological differences between mice and humans, coupled with the fact that mice do not naturally develop PD, have resulted in the failure of many of these mouse-based trials in the end. In this context, Chen et al. (2023) have laid a new foundation for translating cutting-edge biotechnology into feasible pre-clinical practice. Furthermore, recent reports have highlighted various new therapeutic approaches for PD, with several advancing to stages of clinical trial validation. For example, Bayer recently announced positive Phase-I results for its stem cell-based therapy bemdaneprocel (BRT-DA01) for PD, encouraging progression to the next phase of clinical testing. Concurrently, several companies have been developing immunotherapies targeting α-synuclein. Among these, prasinezumab (PRX002), developed by Prothena and Roche, did not achieve the expected outcomes in its Phase-II trials. Nevertheless, PRX002 demonstrated potential signs of disease progression mitigation, such that the companies have resolved to continue a Phase-IIb trial to further validate its efficacy. The ongoing development of novel therapies underscores the complexity of translating basic findings into clinical practice, often presenting more challenges than expected.Given the success of primate PD models, circuit-specific gene therapy shows a promising future in clinical application. Although AAV-mediated therapies are considered safe and sustainable over a long period of time (Kang et al., 2023), challenges remain in terms of potential immune reactions and unknown long-term effects of chemogenetic manipulation. The cost of medication is another concern for stem cell therapies (Chen & Niu, 2019) and immunotherapies. Currently, available immunotherapies (for other diseases) are expensive, hindering expansion of the market. Regarding the therapy proposed by Chen et al. (2023), the market price is uncertain, but it is hoped that it will be affordable for all patients. Although many factors need to be considered before potential clinical trials, we are enthusiastic about the progress made by the team and hope that their therapy opens the door for a new era in PD PETING INTERESTSThe authors declare that they have no competing interests.AUTHORS’ CONTRIBUTIONSY.Q.W., M.Y., and Z.S.G. conceived and wrote the draft. Y.Q.W. generated the figure. All authors read and approved the final version of the manuscript.Yue-Qi Wang1,2,3, Ming Yin1,2,*, Zhe-Shan Guo1,2,*1 State Key Laboratory of Digital Medical Engineering, School ofBiomedical Engineering, Hainan University, Haikou, Hainan570228, China2 Key Laboratory of Biomedical Engineering of Hainan Province, One Health Institute, Hainan University, Haikou, Hainan 570228,China3 Shenzhen Technological Research Center for PrimateTranslational Medicine, Shenzhen-Hong Kong Institute of Brain Science, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, Shenzhen, Guangdong 518055 China *Corresponding authors, E-mail: ********************.cn;**********************.cnREFERENCESBloem BR, Okun MS, Klein C. 2021. Parkinson's disease. The Lancet, 397(10291): 2284−2303.Chen YF, Hong ZX, Wang JY, et al. 2023. Circuit-specific gene therapy reverses core symptoms in a primate Parkinson’s disease model. Cell,doi: 10.1016/j.cell.2023.10.004.Chen ZZ, Niu YY. 2019. Stem cell therapy for Parkinson's disease using non-human primate models. Zoological Research, 40(5): 349−357.Kang L, Jin SL, Wang JY, et al. 2023. AAV vectors applied to the treatment of CNS disorders: clinical status and challenges. Journal of Controlled Release, 355: 458−473.Li H, Su LY, Yang LX, et al. 2021a. A cynomolgus monkey with naturally occurring Parkinson's disease. National Science Review, 8(3): nwaa292.Li H, Yao YG, Hu XT. 2021b. Biological implications and limitations of a cynomolgus monkey with naturally occurring Parkinson's disease. Zoological Research, 42(2): 138−140.Roth BL. 2016. DREADDs for neuroscientists. Neuron, 89(4): 683−694. Volkmann J, Daniels C, Witt K. 2010. Neuropsychiatric effects of subthalamic neurostimulation in Parkinson disease. Nature Reviews Neurology, 6(9): 487−498.Wang R, Shih LC. 2023. Parkinson's disease - current treatment. Current Opinion in Neurology, 36(4): 302−308.Figure 1 Circuit-specific gene therapy opens the door for a new era of Parkinson’s disease (PD) therapeuticsZoological Research 44(6): 1152−1153, 2023 1153。
多巴胺 英语介绍 PPT
dopamine
The discovery of dopamine as a neurotransmitter in brain by Arid Carlsson approximately 50 years ago, and the subsequent insight provided by Paul Greengard into the cellular signalling mechanisms triggered by dopamine, gained these researchers the Nobel Prize for Medicine in 2000.
人的脑中存在着数千亿个神经细胞,人所以有欲望,四肢 躯体灵活运动,都是由于脑部信息在它们之间无阻传递。事 实上,人会有思想,会有感觉,会对一些事物热烈追求,这 都是由于我们大脑内一些微小物质的化学作用。
Dopamine is a natural stimulant. It gives you ecstasy. It gives you focused attention. It gives you motivation and goal-oriented behavior. My hypothesis, which the FMRIs support, is that dopamine is the driving force behind the energy and arousal and attention that are all a part of love.
Arvid Carlsson
保罗· 格林加德(1925-),美国生物医学家, 2000年 获得诺贝尔生理学、医学奖,其在世界上首次提出 了化学因子传导神经细胞信号,这一成果为帕金森病 的治疗与研究作出贡献。
Neuron overload and the juggling physician
Neuron overload and the juggling physicianDanielle Ofri aPatients often complain that their doctors don't listen. Although there are probably a few doctors who truly are tone deaf, most are reasonably empathic human beings, and I wonder why even these doctors seem prey to this criticism. I often wonder whether it is sheer neuron overload on the doctor side that leads to this problem. Sometimes it feels as though my brain is juggling so many competing details, that one stray request from a patient—even one that is quite relevant—might send the delicately balanced three-ring circus tumbling down.One day, I tried to work out how many details a doctor needs to keep spinning in her head in order to do a satisfactory job, by calculating how many thoughts I have to juggle in a typical office visit. Mrs Osorio is a 56-year-old woman in my practice. She is somewhat overweight. She has reasonably well-controlled diabetes and hypertension. Her cholesterol is on the high side but she doesn't take any medications for this. She doesn't exercise as much as she should, and her last DEXA scan showed some thinning of her bones. She describes her life as stressful, although she's been good about keeping her appointments and getting her blood tests. She's generally healthy, someone who'd probably be described as an average patient in a medical practice, not excessively complicated.Here are the thoughts that run through my head as I proceed through our 20-min consultation.Good thing she did her blood tests. Glucose is a little better. Cholesterol isn't great. May need to think about starting a statin. Are her liver enzymes normal?Her weight is a little up. I need to give her my talk about five fruits and vegetables and 30 min of walking each day.Diabetes: how do her morning sugars compare to her evening sugars? Has she spoken with the nutritionist lately? Has she been to the eye doctor? The podiatrist?Her blood pressure is good but not great. Should I add another BP med? Will more pills be confusing? Does the benefit of possible better blood pressure control outweigh the risk of her possibly not taking all of her meds?Her bones are a little thin on the DEXA. Should I start a bisphosphonate that might prevent osteoporosis? But now I'm piling yet another pill onto her, and one that requires detailed instructions. Maybe leave this until next time?How are things at home? Is she experiencing just the usual stress of life, or might there be depression or anxiety disorder lurking? Is there time for the depression questionnaire?Health maintenance: when was her last mammogram? PAP smear? Has she had a colonoscopy since she turned 50? Has she had a tetanus booster in the past 10 years? Does she qualify for a pneumonia vaccine?Ms Osorio interrupts my train of thought to tell me that her back has been aching for the past few months. From her perspective, this is probably the most important item in our visit, but the fact is that she's caught one of my neurons in mid-fire (the one that's thinking about her blood sugar, which is segueing into the neuron that's preparing the diet-and-exercise discussion, which is intersecting with the one that's debating about initiating a statin). My instinct is to put one hand up and keep all interruptions at bay. It's not that I don't want to hear what she has to say, but the sensation that I'm juggling so many thoughts, and need to resolve them all before the clock runs down, that keeps me in moderate state of panic. What if I drop one—what if one of my thoughts evaporates while I address another concern? I'm trying to type as fast as I can, for the very sake of not letting any thoughts escape, but every time I turn to the computer to write, I'm not making eye contact with Mrs Osorio. I don't want my patient to think that the computer is more important than she is, but I have to keep looking toward the screen to get her lab results, check her mammogram report, document the progress of her illnesses, order the tests, refill her prescriptions.Then she pulls a form out her of bag: her insurance company needs this form for some reason or another. An innocent—and completely justified—request, but I feel that this could be the straw that breaks the camel's back, that the precarious balance of all that I'm keeping in the air will be simply unhinged. I nod, but indicate that we need to do her physical examination first. I barrel through the basics, then quickly check for any red-flag signs that might suggest that her back pain is anything more than routine muscle strain. I return to the computer to input all the information, mentally running through my checklist, anxious that nothing important slips from my brain's holding bay.I want to do everything properly and cover all our bases, but the more effort I place into accurate and thorough documentation, the less time I have to actually interact with my patient. A glance at the clock tells me that we've gone well beyond our allotted time. I stand up and hand Mrs Os orio her prescriptions. “What about my insurance form,” she asks. “It needs to be in by Friday, otherwise I might lose my coverage.” I clap my hand against my forehead; I've completely forgotten about the form she'd asked about just a few minutes ago.Studies have debunked the myth of multitasking in human beings. The concept of multitasking was developed in the computer field to explain the idea of a microprocessor doing two jobs at one time. It turns out that microprocessors are in fact linear, and actually perform only one task at a time. Our computers give the illusion of simultaneous action based on the microprocessor “scheduling” competing activities in a complicated integratedalgorithm. Like microprocessors, we humans can't actually concentrate on two thoughts at the same exact time. We merely zip back and forth between them, generally losing accuracy in the process. At best, we can juggle only a handful of thoughts in this manner. The more thoughts we juggle, the less we are able to attune fully to any given thought. To me, this is a recipe for disaster. Today I only forgot an insurance company form. But what if I'd forgotten to order her mammogram, or what if I'd refilled only five of her six medicines? What if I'd forgotten to fully explain the side-effects of one of her medications? The list goes on, as does the anxiety.At the end of the day, my mind spins as I try to remember if I've forgotten anything. Mrs Osorio had seven medical issues to consider, each of which required at least five separate thoughts: that's 35 thoughts. I saw ten patients that afternoon: that's 350. I'd supervised five residents that morning, each of whom saw four patients, each of whom generated at least ten thoughts. That's another 200 thoughts. It's not to say that we can't handle 550 thoughts in a working day, but each of these thoughts potentially carries great risk if improperly evaluated. If I do a good job juggling 98% of the time, that still leaves ten thoughts that might get lost in the process. Any one of those lost thoughts could translate into a disastrous outcome, not to mention a possible lawsuit. Most doctors are reasonably competent, caring individuals, but the overwhelming swirl of thoughts that we must keep track of leaves many of us in a perpetual panic that something serious might slip. This is what keeps us awake at night.There are many proposed solutions—computer-generated reminders, case managers, ancillary services. To me, the simplest one would be time. If I had an hour for each patient, I'd be a spectacular doctor. If I could let my thoughts roll linearly and singularly, rather than simultaneously and haphazardly, I wouldn't fear losing anything. I suspect that it would actually be more efficient, as my patients probably wouldn't have to return as frequently. But realistically, no one is going to hand me a golden hour for each of my patients. My choices seem to boil down to entertaining fewer thoughts, accepting decreased accuracy for each thought, giving up on thorough documentation, or having a constant headache from neuronal overload.These are the choices that practising physicians face every day, with every patient. Mostly we rely on our clinical judgment to prioritise, accepting the trade-off that is inevitable with any compromise. We attend to the medical issues that carry the greatest weight and then have to let some of the lesser ones slide, with the hope that none of these seemingly lesser ones masks something grave.Some computers have indeed achieved the goal of true multitasking, by virtue of having more than one microprocessor. In practice, that is like possessing an additional brain that can function independently and thus truly simultaneously. Unless the transplant field advances drastically, there is little hope for that particular deus ex machina. In some cases,having a dedicated and competent clinical partner such as a one-on-one nurse can come close to simulating a second brain, but most medical budgets don't allow for such staffing indulgence.As it stands, it seems that we will simply have to continue this impossible mental high-wire act, juggling dozens of clinical issues in our brains, panicking about dropping a critical one. The resultant neuronal overload will continue to present a distracted air to our patients that may be interpreted as us not listening, or perhaps not caring.When my computer becomes overloaded, it simply crashes. Usually, I reboot in a fury, angry about all my lost work. Now, however, I view my computer with a tinge of envy. It has the luxury of being able to crash, and of a reassuring, omniscient hand to press the reboot button. Physicians are permitted no such extravagance. I pull out the bottle of paracetamol tablets from my desk drawer and set about disabling the childproof cap. It's about the only thing I truly have control over.。
神经生理学的英文名词解释
神经生理学的英文名词解释神经生理学是研究神经系统的功能和活动的学科,结合生物学和生理学的知识,涵盖了神经元、神经网络以及神经传递的研究。
以下将介绍一些与神经生理学相关的英文名词,为读者提供更深入的了解。
1. Neuron(神经元)Neurons, also known as nerve cells, are the fundamental building blocks of the nervous system. They are specialized cells that transmit electrical signals throughout the body. Neurons have three main components: the cell body, dendrites, and axons. The cell body contains the nucleus and other organelles, while the dendrites receive signals from other neurons. The axon carries electrical impulses away from the cell body to other neurons or target cells.2. Synapse(突触)A synapse is a junction between two neurons where communication occurs. It is the site where the electrical signal from one neuron is transmitted to another neuron or target cell. Synapses can be either electrical or chemical. In chemical synapses, neurotransmitters are released from the presynaptic neuron and bind to receptors on the postsynaptic neuron, allowing the electrical signal to be transmitted.3. Action potential(动作电位)An action potential is a rapid and brief change in the membrane potential of a neuron. It is the result of depolarization and repolarization of the neuron's membrane, which allows the electrical signal to be conducted along the axon. The action potential is an "all-or-nothing" response, meaning it either occurs fully or not at all.4. Neurotransmitter(神经递质)Neurotransmitters are chemical messengers that transmit signals across synapses. They are released from the presynaptic neuron and bind to receptors on the postsynapticneuron, causing changes in the electrical activity of the postsynaptic neuron. Examples of neurotransmitters include dopamine, serotonin, and acetylcholine.5. Central nervous system(中枢神经系统)The central nervous system (CNS) consists of the brain and spinal cord. It is responsible for processing and integrating information received from sensory neurons and initiating appropriate responses. The CNS controls various functions, including movement, learning, memory, and emotion.6. Peripheral nervous system(外周神经系统)The peripheral nervous system (PNS) includes all the nerves and ganglia outside of the CNS. It connects the CNS to the rest of the body and is responsible for transmitting sensory information to the CNS and motor commands from the CNS to muscles and organs. The PNS can be further divided into the somatic nervous system and the autonomic nervous system.7. Somatic nervous system(躯体神经系统)The somatic nervous system controls voluntary movements and transmits sensory information from the body to the CNS. It consists of sensory neurons, motor neurons, and interneurons. Motor neurons carry signals from the CNS to muscles, allowing for voluntary movements.8. Autonomic nervous system(自主神经系统)The autonomic nervous system regulates involuntary processes in the body, such as heartbeat, digestion, and breathing. It has two main divisions: the sympathetic nervous system and the parasympathetic nervous system. The sympathetic nervous system prepares the body for "fight-or-flight" responses, while the parasympathetic nervous system promotes "rest-and-digest" activities.通过以上对于神经生理学的英文名词解释,我们可以更好地理解神经系统的基本原理和功能。
袁俊英细胞凋亡综述Apoptotic and nonapoptotic roles of caspases
Caspases, a family of cysteine proteases, are crucial intracellular signal transducers and executioners of apoptosis , which is a genetically encoded cellular sui-cide pathway. Caspases may be activated in response to different pro-death stimuli through the formation of activating complexes in the presence of adaptor pro-teins. Activated caspases proteolytically cleave both additional caspases and specific cellular substrates to cause cellular destruction during apoptosis. Activation of caspases has been implicated in mediating neuronal cell death during development as well as neuronal loss in both acute and chronic neurodegenerative diseases. Caspases are also important mediators of inflammatory responses. Activation of specific caspases can lead to the processing of pro-interleukin-1β and the produc-tion of mature interleukin-1β, which is an important pro-inflammatory cytokine.Unexpectedly, recent studies show that caspases also mediate regulatory events in the mature nerv-ous system that are crucial for neural functions, such as axon pruning and synapse elimination, which are important to refine mature neuronal circuits. Thus, although a global activation of caspases leads to apo-ptosis, restricted and localized activation may control normal physiology in living neurons. This Review explores the multiple faces of caspase activity in regu-lating both apoptotic and non-apoptotic processes in neurons.Discovery of the key mediators of apoptosisStudies of programmed cell death in the nematode Caenorhabditis elegans provided the first insights into the mechanism of apoptosis in mammalian cells. During development, C. elegans produces 1,090 somatic cells (of which 302 become neurons and 56 become glial cells) in each young adult worm; an additional 131 cells are generated but subsequently undergo programmed cell death 1. The fate of these 131 cells is predetermined by their genetic lineages and occurs at predictable times during the development of each worm. In the same manner, the surviving 1,090 somatic cells are predeter-mined by their genetic lineages to develop into specific cells such as neurons or muscle. Thus, the cell death that occurs in C. elegans during development is a form of genetically programmed cellular suicide.By contrast, there is no evidence to indicate that neu-ronal cell death during the development of the mam-malian nervous system is predetermined by cell lineage. Instead, the decision of whether and which neurons will live or die is reached through a competitive process during which immature neurons compete for appropri-ate innervation targets and supplies of trophic factors 2. Therefore, neuronal cell death during mammalian devel-opment is a stochastic process that is not predetermined by cell lineage. Interestingly and unexpectedly, however, the stochastic neuronal cell death that occurs during the development of higher organisms shares mechanistic1Neurology Service, Massachusetts General Hospital, 114 16th Street Charlestown, Massachusetts 01029, USA.2Department of Cell Biology, Harvard Medical School, 240 Longwood Avenue, Boston, Massachusetts 02115, USA.e-mails:bhyman@ ; jyuan@ doi:10.1038/nrn3228ApoptosisA form of regulated cell death that is controlled by themembers of the caspase family and BCL‑2 family of proteins.Programmed cell deathCell death that is regulated by genetically encoded mechanisms.Apoptotic and non-apoptotic roles of caspases in neuronal physiology and pathophysiologyBradley T . Hyman 1 and Junying Yuan 2Abstract | Caspases are cysteine proteases that mediate apoptosis, which is a form of regulated cell death that effectively and efficiently removes extra and unnecessary cells duringdevelopment. In the mature nervous system, caspases are not only involved in mediating cell death but also regulatory events that are important for neural functions, such as axon pruning and synapse elimination, which are necessary to refine mature neuronal circuits. Furthermore, caspases can be reactivated to cause cell death as well as non-lethal changes in neurons during numerous pathological processes. Thus, although a global activation ofcaspases leads to apoptosis, restricted and localized activation may control normal physiology and pathophysiology in living neurons. This Review explores the multiple roles of caspase activity in neurons.The members of the BCL‑2 family share one or more of the four BH domains (BH1, BH2, BH3 and BH4) that mediate the interactions among different members of theBCL‑2 family.Caspase recruitment domain(CARD). A protein–protein interacting motif frequently found in proteins associated with apoptosis and inflammation.Death effector domain (DED). A protein–protein interacting motif found in proteins that mediate caspase activation and apoptosis. Death-inducing signalling complex(DISC). Intracellular signalling complexes associated with the receptors in the tumour necrosis receptor 1 superfamily that are involved in mediating cell death.. This Review examines the uniqueelucidated roles for these same pathways in normal neu-ronal function as well as in pathophysiological events.Four genes control the programmed cell deathmachinery in C. elegans : egl‑1, ced‑3, ced‑4 and ced‑9.egl‑1 and ced‑9 are pro-apoptotic and anti-apoptoticmembers, respectively, of the bcl‑2 (B-cell leukemia/lymphoma 2) family of proteins; ced‑3 encodes a mem-ber of the caspase family; and ced‑4 encodes an activatorof ced‑3. In mammals, the homologues of EGL-1 andCED-9 are also members of the BCL-2 family of pro-teins, the homologue of CED-3 also encodes caspases,and the homologue of CED-4 is apoptotic peptidase-activating factor 1 (APAF1; a member of the nucleo-tide-binding and oligomerization domain (NOD)-likereceptors (NLRs)). When a cell undergoes programmedcell death during C. elegans development, the expres-sion of egl‑1 may be induced. The function of EGL-1 isto displace CED-9 from its complex with CED-4, and,once released, CED-4 interacts with CED-3 to mediatethe proteolytic activation of CED-3. The first insightsinto a role for apoptosis in regulating neuronal celldeath was provided by studies showing that neuronalcell death induced by the lack of trophic factors can beprevented by overexpression of BCL-2 or by expressionof a virally encoded caspase inhibitor crmA3. Thesefindings suggest that the stochastic neuronal cell deathprocess that occurs during mammalian development isregulated by a cellular suicide machinery that is likelyto share the same evolutionary origin as that of pro-grammed cell death in C. elegans, which is regulated bythe cell lineage.CED-9 and mammalian anti-apoptotic BCL-2 fam-ily members, such as BCL-2, BCL-XL(also known asBCL2L1), MCL1 (myeloid cell leukemia sequence 1(BCL2-related)) and BCL-w (also known as BCL2L2),contain multiple BCL‑2 homology domains (BH1, BH2,BH3 and BH4) that mediate the heterodimeric inter-action with pro-apoptotic members of the BCL-2 fam-ily. They also contain a carboxy-terminal hydrophobictransmembrane tail domain that is involved in anchoringthe protein to the outer mitochondrial membrane.Members of the pro-apoptotic BCL-2 family of pro-teins, however, are divided into two subgroups: thosethat contain multiple BH domains (for example, BCL-2-associated X protein (BAX) and BCL-2-antagonist/killer(BAK)) and those that have only the BH3 domain (forexample, BH3 interacting domain death agonist (BID),BIM (also known as BCL2L11), PUMA (also knownas BBC3) and BCL-2-associated agonist of cell death(BAD)). During apoptosis, BAX and BAK oligomerizeand insert into the outer mitochondrial membrane tomediate mitochondrial damage and cytochrome c release,which leads to the activation of caspases4.Two models have been proposed to explain the acti-vation of BAX and BAK during apoptosis: through director indirect interactions with BH3-only pro-apoptoticBCL-2 family members. In the direct model, BH3-onlypro-apoptotic proteins are divided into activators, suchas BID, BIM, PUMA, and sensitizers, such as BAD. Theactivators directly interact with and induce conforma-tional changes in BAK and BAX5, whereas sensitizersbind to anti-apoptotic proteins to stimulate release ofactivator BH3-only proteins, therefore leading to theactivation of BAK and BAX. The anti-apoptotic BCL-2family members, such as BCL-2 and BCL-XL, bind andsequester BH3-only proteins, and can also bind to acti-vated BAK and BAX proteins to prevent apoptosis. Inthe indirect model, BCL-2 and BCL-XLalways bind toBAK and BAX, which blocks their activation. The releaseand subsequent activation of BAX and BAK is causedby death signals that induce the binding of BH3-onlyproteins to anti-apoptotic proteins6.Involvement of caspases in apoptosis. In contrast to pro-grammed cell death in C. elegans in which CED-3 is theonly caspase required, multiple caspases are involvedin mediating apoptosis in higher organisms; for exam-ple, the mouse genome encodes 10 caspases, whereasthe human genome encodes 11 caspases. Membersof the caspase family are subdivided into upstream initia-tor caspases (including caspases 1, 2, 4, 5, 8, 9, 10, 11 and12), which respond to pro-apoptotic signals and activatedownstream executioner caspases (including caspases3, 6, 7 and 14). Initiator caspases possess a long amino-terminal prodomain, which mediates the interaction ofother proteins in complexes to then interact with theiractivators and inhibitors. However, not all caspases areprimarily involved in apoptosis, as some are involvedin diverse biological processes that include immuneregulation and spermatogenesis7.Caspases are synthesized as zymogens comprisinga prodomain, a large subunit and a small subunit. Theactivation of caspases requires allosteric conformationalchanges and/or specific cleavage after a specific aspar-tate residue and, thus, the cleavage of caspases has beenrecognized as a hallmark of caspase activation. Theinitial processing of the inactive caspase separates thelarge and small subunits, and the subsequent removalof the N-terminal prodomain then activates the pro-tease8. Apoptosis is a culmination of a signalling cascadewhereby activated upstream caspases cleave downstreamcaspases and their respective substrates.The activation of upstream caspases is a crucialevent during apoptosis (FIG. 1). The long prodomainsof upstream caspases are involved in mediating theirinteractions with their respective adaptor proteins. Thelong prodomain of caspases exists in two forms: thecaspase‑recruitment domain (CARD) or the death‑effectordomain (DED). Adaptor proteins recruit specific initia-tor caspases through a homotypic interaction with thisdomain8. Four specific caspase-activating complexeshave been characterized in mammals: the apoptosome,the death‑inducing signalling complex (DISC), the inflam-masomes and the the PIDDosome. The apoptosomemediates the activation of caspase 9 by interacting withthe adaptor APAF1 in the presence of cytochrome cin the mitochondrial pathway of apoptosis. DISC medi-ates the activation of caspase 8 by interacting with theNature Reviews Neuroscienceadaptor FADD in the death receptor-induced apoptosis pathway. The inflammasomes mediate the activation of caspase 1 and caspase 5 in inflammatory responses by interacting with the adaptors ASC (apoptosis-associated speck-like protein containing a CARD) or the family of NLRs. The PIDDosome mediates the activation of cas-pase 2 by interacting with the adaptors RAIDD (receptor- interacting protein (RIP)-associated ICH-1/CED-3 homologous protein with a death domain; also called CRADD)9 and PIDD (p53-induced protein with a death domain) in the DNA damage-induced apoptosis path-way10. The presence of four different caspase-activating complexes presumably allows fine and highly specialized control of apoptosis so that only specific caspases are activated in response to appropriate pro-apoptotic and pro-inflammatory signals.Figure 1 | Caspase activation occurs in response to various stimuli. Downstream caspases (including caspase 3) are activated in several specialized scenarios, and four are outlined here: the classical apoptosome, in which the activation of caspase 3 occurs by intrinsic, mitochondrial mechanisms (a), and extrinsic cell surface receptor-mediated processes (b) following inflammatory stimuli (c) or after DNA damage (d). In each scenario, upstream mediators of caspase activation lead to activation of caspases 1, 2, 8 or 9, which, in turn, activate caspase 3 by cleaving inactive procaspase 3 to its active dimeric form. Classically, caspase 3 is viewed as an executioner caspase as the activation of this protease leads to apoptosis. Crosstalk across these pathways, mitochondrial signals, as well as other stimuli, including endoplasmic reticulum stress and calcium flux, can also initiate or accelerate these programmes of caspase activation. APAF1, apoptotic peptidase-activating factor 1; ASC, apoptosis-associated speck-like protein containing a CARD; BID, BH3 interacting domain death agonist; CARD, caspase-recruitment domain; DISC, death-inducing signalling complex; FADD, FAS-associated death domain protein; FASL, FAS ligand; FIIND, function-to-find domain; LRR, leucine-rich repeat; NLRP, NOD-, LRR- and pyrin domain-containing; PIDD, p53-induced protein with a death domain; PYD, pyrin domain; RAIDD, receptor-interacting protein (RIP)-associated ICH-1/CED-3 homologous protein with a death domain; also called CRADD.Caspase activation in neurons during development The activation of caspases during neuronal cell death is regulated by members of the BCL-2 family. The anti-apoptotic members of the BCL-2 family, which include BCL-2, BCL-X L , MCL1 and BCL-w, play important roles in maintaining neuronal survival during embryonic development and in adult life.Overexpression of BCL-2 prevents apoptosis of sympathetic neurons that is induced by nerve growth factor (NGF) deprivation in cell culture 3. Moreover, transgenic mice overexpressing BCL-2 in the nervous system show reduced neuronal cell death and increased numbers of facial motor neurons and retinal ganglion cells 11. Overexpression of BCL-2 in mice also results in larger brains with more neurons, but otherwise nor-mal brain morphology. By contrast, deficiency of pro-apoptotic caspase 9 or APAF1 leads to severe defects in brain morphogenesis (discussed further below). Given the brain morphology of BCL-2 transgenic mice, one might argue that a simple reduction in neuronal cell death is not sufficient to significantly alter brain mor-phogenesis. However, the more severe consequences of caspase 9 and APAF1 deficiency suggest that these proteins might have additional functions unrelated to regulation of apoptosis, a possibility that should be examined in the future.The expression of BCL-2 is high in the CNS during development and downregulated after birth, eventually becoming negligible except in peripheral neurons where it remains at high levels throughout life. Mice deficient in BCL-2 (Bcl2–/–) have normal prenatal nervous system development, but after birth there is a loss of motor, sensory and sympathetic neurons, which suggests that BCL-2 is important for maintaining specific populations of neurons at this time 12.The apparently normal prenatal nervous sys-tem development of Bcl2–/– mice might be because of functional redundancy within the BCL-2 family. The expression of BCL-X L (which is also expressed in the developing brain) decreases after birth but remains at detectable levels in the adult CNS 13. In contrast to Bcl2–/– mice, Bclxl –/– mice die around embryonic day 13, with extensive apoptotic cell death in postmitotic differenti-ating neurons of the developing brain, spinal cord and dorsal root ganglia. This striking deficiency of neuronal survival in Bclxl –/– mice indicates that BCL-X L has a pivotal role in maintaining neuronal survival.The levels of MCL1, an anti-apoptotic BCL-2 homo-logue, are rapidly regulated following cellular stress through both synthesis and degradation pathways 14. Mcl1–/– mice die early in embryonic development (at around embryonic day 3.5), which is the most severe phenotype among all the known transgenic mice bear-ing mutations in the members of the anti-apoptotic BCL-2 family. Robust expression of MCL1 is present in both proliferating neuronal progenitor cells and in postmitotic neurons during brain development. MCL1 deficiency, especially in the neuronal lineage, results in the apoptotic death of both nestin-positive neural pro-genitors and TUJ1-positive newly committed neurons. During the development of the nervous system in MCL1conditional mutants, pro-apoptotic BAX and BID are expressed in neural precursor cells within the ventricu-lar zone, with the expression of BAX peaking at embry-onic days 12 to 15. This corresponds to the period of early neurogenesis when widespread apoptosis of neural precursors occurs. Although both MCL1 and BCL-X L have been shown to block BAX- and BAK-mediated apoptosis, BCL-X L is expressed at very low levels in the neural precursor populations, and apoptosis is observed in more mature neuronal populations in the Bclxl –/– mice at embryonic day 12.5. Together, these data suggest that MCL1 may play an important role in regulating the survival of neurons during the transition from the progenitor to the postmitotic period.The expression of another member of the anti-apoptotic BCL-2 family, BCL-w, is selectively induced in the axons of sensory neurons in response to target-derived neurotrophin stimulation, which suggests that BCL-w may be involved in promoting the survival of neurons that can establish appropriate synaptic connec-tions 15. Bclw –/– mice exhibit a progressive and concurrent degeneration of small fibre innervation in the skin as well as in the ability to detect noxious thermal stimuli. Cell body number is unaffected at symptomatic ages, which suggests that this degeneration is initiated in the axons. Furthermore, the axonal mitochondria in sensory neurons of Bclw –/– mice exhibit abnormalities in size and function, which is consistent with the known function of BCL-w in maintaining the mitochondrial functions in axons.Wild-type neonatal sympathetic superior cervical ganglion neurons and facial motor neurons express Bax mRNA at a time when these neuronal populations are susceptible to growth factor deprivation in vivo . Deletion of Bax results in profound effects on the survival of many kinds of neurons. Neonatal sympathetic neurons defi-cient in BAX survive NGF deprivation in culture, and facial motor neurons deficient in BAX survive discon-nection from their targets by axotomy in vivo 16. Bax –/– mice have increased neuronal numbers in the superior cervical ganglia and facial nuclei. Therefore the activa-tion of BAX may be a crucial event in initiating neuronal cell death both during the development of these regions at times of trophic factor withdrawal, as well as follow-ing injury. Although there is a great excess of neurons in postnatal Bax –/– mice, only a subpopulation of axoto-mized motor neurons rescued with exogenous trophic factor can grow and retain their muscle innervation target contacts postnatally. Such treatment can reverse axonal atrophy and promote regrowth of the axons of the excess surviving motor neurons, which suggests that even after their developmental role in regulating motor neuron survival is complete, trophic signals are still necessary to promote cell growth and to sustain target innervation.Role of apoptosis in the mature nervous system. The BCL-2 family of proteins is also involved in mediating neurogenesis in the adult nervous system. The olfactory bulb is one of the few brain regions in which neurogen-esis continues throughout adult life. The neural stemExencephalyA type of developmental abnormality whereby the brain may protrude outside the skull.cells that produce the olfactory interneurons in the adultbrain mainly reside in a specialized cellular niche locatedin the anterior subventricular zone of the lateral ven-tricle. Approximately 30–70% of newly generated neu-rons in the adult brain undergo programmed cell deathwithin 1 month of their initial production. Programmedcell death of adult-generated neurons involves BAX. Anincreased number of new neurons in the subcallosalzone is found in Bax–/–mice17. Treatment with either acaspase inhibitor or with the neurotrophic factor BDNF(brain-derived neurotrophic factor) promoted the sur-vival of adult-generated neurons in adult wild-type mice,which suggests that insufficient neurotrophic signallingin the subcallosal zone triggers BAX-dependent apop-tosis of newly generated neurons in this area. However,most surviving adult-produced cells in the subcallosalzone of Bax–/– mice fail to differentiate into normalneuronal nuclei-positive neurons and undergo atrophy.Thus, blocking apoptosis may not be sufficient to allowthe maturation of newly generated neurons.The activation of BAX and/or BAK during neu-ronal apoptosis is mediated by the BH3-only mem-bers of BCL-2 family of proteins, which are induced inresponse to trophic factor deprivation. However, there isfunctional redundancy among the various members ofthe BH3-only proteins. Although deficiency in indi-vidual BH3-only genes can reduce or delay neuronalcell death, only simultaneous deficiency of three of theBH3-only genes, Bid, Bim and Puma, can block the deathof cerebellar granule neurons that is induced by DNAdamage and potassium deprivation as strongly as thatof neurons deficient in both BAX and BAK18. Theseresults suggests that BID, BIM and PUMA work togetherto mediate the activation of BAX and BAK in neuronsunder these experimental conditions.Roles of caspase-induced neuronal cell deathThere is considerable evidence to suggest that caspasecleavage initiates apoptosis. Neuronal cell death that isinduced by trophic factor deprivation in vitro can beblocked by the expression of viral inhibitors of upstreamcaspases (such as those encoding crmA)3 or by peptide-based chemical inhibitors19. The cleavage of downstreamcaspases, such as caspase 3 and caspase 7, is also associ-ated with neuronal cell death in both the developing andmature nervous systems. However, mammalian genomesencode multiple functional caspases, and this has madeit more challenging to elucidate the precise in vivo rolesof caspases in regulating neuronal cell death during pro-cesses such as synaptogenesis, which varies according tobrain region and species.Although mutant mice for all caspases have been gen-erated, none of the caspase-deficient mice has shown asignificant defect in the elimination of mature neurons,which may be due to the presence of alternative celldeath pathways. The basis of this apparent redundancyis still largely unexplored, but it implies that there aremultiple levels of caspase activation and regulation inmammalian systems. Drosophila melanogaster modelshave also been used in attempts to investigate the func-tion of individual caspases in neuronal cell death, butapoptotic cells in this species seem to generate meitoticsignals, which adds an additional level of complexity.The extent to which this occurs in mammalian systemsis less well understood20.Mouse mutants for caspase 3, caspase 9 and APAF1exhibit an increase in the thickness and surface of thecerebral cortex, a drastic reduction in early cerebralapoptosis and an inappropriate amplification of spe-cific populations of neural progenitors, which all resultin brain exencephaly and neural overgrowth21–25. Theseobervations suggest that during early development of theforebrain, caspase 3, caspase 9 and APAF1 are involvedin controlling the numbers of mitotically active neuralprogenitor cells and immature neurons by apoptosis.Conversely, however, significant apoptosis was detectedin the proliferative zone of the mouse brain.The mouse strain used can also influence the pheno-typic outcome. For instance, mice deficient in caspase 3(Casp3–/–) bred from a mixed 129/SvJ and C57BL/6background die perinatally and show various hyper-plasias and disorganized cell migration in the brain,similar to that observed in Casp9−/− mice. Ectopic cellmasses appear in the cerebral cortex, hippocampus andstriatum, whereas the incidence of pyknosis (shrinkingof the nucleus accompanied by chromatin condensa-tion), a prominent feature of apoptosis during normalneurogenesis in the periventricular zone, is considerablyreduced in Casp3−/− and Casp9−/− mice. The phenotypesof Casp3−/− mice are strongly dependent on the geneticbackground, as Casp3−/− mice are viable and develop-mentally normal in C57BL/6 background. These resultssuggest that the capacity of genetic compensation forcaspase 3 deficiency is strain-specific. The presenceof ectopic neuronal nests is reminiscent of alterationsseen in some neurodevelopmental diseases, and also ofthe abnormal cytoarchtiecture reported in some casesof schizophrenia26, thus raising the possibility that theneurodevelopmental processes in which caspases play apart may be involved in these illnesses.Neural precursor cells in the periventricular zoneof Casp3−/− and Casp9−/− mice exhibit apoptotic defectsbefore they migrate to their target locations and formsynaptic connections. Target-dependent, naturallyoccurring neuronal death that occurs when synapticconnections are being formed is more well known andmay differ from the type that occurs before the forma-tion of synaptic connections. Interestingly, in contrast tothe striking perturbations in the morphology observedin the more rostral, mitotically active regions of thebrains of Casp3−/− and Casp9−/− embryos, the spinal cord,brainstem and peripheral ganglia in these mice appearcompletely normal at both embryonic and postnatalstages. Developing postmitotic neurons in these regionsultimately undergo normal neuronal loss, despite thetemporal delay in cell death27. Thus, although the stud-ies using in vitro models of trophic factor-dependentneuronal cell death implicate an important role of cas-pases in apoptosis, there is a general lack of evidence forthe requirement of any specific caspase in postmitoticneuronal death in vivo. There are several reasons thatmight explain this.NecroptosisA necrotic cell death mechanism regulated by the RIP1 and RIP3 kinases.It is possible that the constitutive loss of a caspase inthe germ line might lead to a compensatory upregula-tion of the expression of other caspases in the mutantbackground. This has been demonstrated for Casp3−/−,Casp7−/− and Casp9−/− mice. However, compensatoryexpression of other caspases does not explain the lackof obvious deficiency in neuronal cell death in Casp3−/−Casp7−/− double-knockout embryos, which lack both ofthe major downstream caspases necessary for the execu-tion of apoptosis28. To rule out possible compensatoryexpression of different caspases that could explain thelack of defects in neuronal cell death in these caspasedeficient mice, it will be interesting to examine whetherconditional caspase mutant mice that are specificallydeficient for a caspase in the neuronal lineage or in atemporally regulated manner show a defect in neuronalcell death that is induced by trophic-factor deprivation.A second reason for the lower than expected level of apo-ptotic cell death in these caspase-deficient mice is that anadditional, caspase-independent, cell death mechanismis activated to compensate for the caspase deficiency.In support of this notion, although many populationsof developing postmitotic neurons exhibit normalamounts of cell death in the absence of either caspase 3or caspase 9, both the kinetics of degradation and themorphological changes observed in these degeneratingneurons differ from the more typical apoptotic cell death.Ultrastructural analysis of degenerating spinal cord neu-rons from Casp3−/− embryos at day 14.5 revealed the pres-ence of extensive cytoplasmic vacuoles that are indicativeof cell death by a non-apoptotic, caspase-independentpathway. These vacuoles are not usually observed in apo-ptotic cells or in dying neurons of wild-type embryos. Adelay in the degeneration of these caspase-deficient neu-rons suggests that caspase-mediated neuronal cell deathis more efficient than caspase-independent cell death.Similar to Casp9−/− mice, regardless of genetic back-ground, Apaf1−/− mice also display exencephaly, whichwas attributed partly to the reduced cell death of bothimmature neurons and dividing neuronal precursor cells.This is also consistent with the role of APAF1 as an acti-vator of caspase 9. In mature neurons, however, APAF1is dispensable for neuronal cell death caused by the lackof trophic signalling input from TRKA deficiency or syn-aptic activity from MUNC18 deficiency in vivo29. Manypopulations of postmitotic, target-dependent neurons(including spinal and cranial motor neurons, spinalinterneurons, dorsal root ganglion sensory neurons andsympathetic neurons in the superior cervical ganglion)undergo a quantitatively normal amount of cell death inthe absence of APAF1. The degenerating Apaf1−/− neu-rons show numerous vacuoles atypical of apoptosis,which suggests that there is activation of an alternativecell death mechanism in developing neurons when cas-pase activation is inhibited. Thus, it is possible that inpostmitotic mature neurons in which caspase activationis impaired, a lack of trophic factor support might trig-ger an active form of cell death that is mediated througha caspase-independent mechanism. It will be interestingto examine whether necroptosis is responsible for thiscaspase-independent neuronal cell death30,31.Non-apoptotic roles of caspasesDendritic pruning. Dynamic pruning of axons and den-drites in the nervous system is important for normalbrain functions. Such pruning can lead to the removalof aberrant or misguided axon branches in the absence ofcell death. The elimination of axon branches can involvesmall-scale pruning of axon terminals at the single-celllevel or a large-scale removal of inappropriate collateralbranches. Local activation of caspases — for example, byinteraction with toxic domains of the amyloid precursorprotein (APP) — can lead to axonal regression32.Caspases have been implicated in influencing neuralconnectivity by regulating dendritic pruning, a processthat involves the selective removal of neuronal dendrites,and it is through this mechanism that caspases are likelyto play a substantial role in regulating the structure andthe function of the CNS. Dendritic pruning is a multistepprocess33,34 involving cytoskeletal destabilization, whichleads to the severing of dendrites from the soma andsubsequent engulfment by glia. Thus, selective removalof certain dendrites might be a form of subcellular apo-ptosis. Consistent with this proposal, localized caspaseactivity is important for the dendritic pruning of sen-sory neurons during fly metamorphosis33,34. Dronc (alsoknown as Nedd2-like caspase; Nc), a major upstreamcaspase in D. melanogaster, is activated in a localizedmanner in the dendrites, which suggests that Dronc isinvolved in the initiation of dendritic pruning during flylarval development. The activation of Dronc is mediatedby the degradation of DIAP1 (Drosophila inhibitor ofapoptosis protein; also known as th), a potent inhibi-tor of caspases, by the ubiquitin–proteasome system.DIAP1 is an E3 ubiquitin ligase that targets Dronc andthus targets it for degradation through the ubiquitin–proteasome system33. Thus, spatially restricted degrada-tion of DIAP1 might control caspase activation locally35.Whether ubiquitin-dependent degradation of DIAP1 isinitiated in a restricted manner in selected dendritesremains unknown, but it is a plausible model for thistype of selective, anatomically localized event.The functional roles for caspase-mediated dendriticpruning are beginning to be elucidated. For example,activation of caspase 3 may play a role in a zebra finchmodel of learning and memory36. Caspase 3 is activatedbriefly in the dendritic spines of zebra finch auditoryforebrain following stimuli associated with memoryformation (neural activation and subsequent responsehabituation). In a similar manner to the regulation ofcaspases in fly neurons by DIAP1 as discussed above, theactivated caspase 3 protein is present even in unstimu-lated brains but bound to BIRC4 (also known as XIAP),a homologue of DIAP and a member of inhibitor of apo-ptosis (IAP) family of caspase inhibitors, thereby sug-gesting an evolutionarily conserved mechanism for rapidrelease and sequestration at specific synaptic sites. Thus,caspase activity may influence postsynaptic remodellingin response to certain environmental stimuli.Axon guidance and synaptogenesis. Caspase-deficientmice exhibit certain specific defects in axon targeting andsynapse formation during development. Considerable。
of
Information in the first spike,the order of spikes,and the number of spikes provided by neurons in the inferior temporal visual cortexEdmund T.Rolls *,Leonardo Franco 1,Nikolaos C.Aggelopoulos,Jose M.Jerez1University of Oxford,Department of Experimental Psychology,South Parks Road,Oxford OX13UD,UKReceived 26October 2005;received in revised form 18July 2006AbstractInformation theoretic analyses showed that for single inferior temporal neurons and neuronal populations,more information was encoded in 20or more ms by all the spikes available than just by the first spike in the same time window about which of 20objects or faces was shown.Further,the temporal order in which the first spike arrived from different simultaneously recorded neurons did not encode more information than was present in the first spike or the spike counts.Thus information transmission in the inferior tem-poral cortex by the number of spikes in even short time windows is fast,and provides more information than only the first spike,or the spike order from different neurons.Ó2006Elsevier Ltd.All rights reserved.Keywords:Object recognition;Inferior temporal cortex;First spike;Spike order1.IntroductionThe question of how information is encoded by neuro-nal activity in the brain is fundamental for understanding how the brain operates.Towards the end of the primate ventral visual system,in the inferior temporal visual cortex,neurons respond with some selectivity to different faces or objects (Baddeley et al.,1997;Desimone,1991;Perrett,Rolls,&Caan,1982;Rolls,2000,2005b;Rolls &Deco,2002;Rolls &Tovee,1995;Rolls,Treves,Tovee,&Panzeri,1997b;Tanaka,1996;Treves,Panzeri,Rolls,Booth,&Wakeman,1999).An important issue is how rapidly information can be read from these neurons,or from any stage of visual cortical processing.A rapid readout of information from any one stage is important,for the ven-tral visual system is organised as a hierarchy of cortical areas,and the neuronal response latencies are approxi-mately 100ms in the inferior temporal visual cortex,and 40–50ms in the primary visual cortex,allowing only approximately 50–60ms of processing time for V1–V2–V4–inferior temporal cortex (Baylis,Rolls,&Leonard,1987;Nowak &Bullier,1997;Rolls &Deco,2002).There is much evidence that the time required for each stage of processing is relatively short.For example,visual stimuli presented in succession approximately 15ms apart can be separately identified (Keysers &Perrett,2002);the reaction time for identifying visual stimuli is relatively short (Bacon-Mace,Mace,Fabre-Thorpe,&Thorpe,2005;Thorpe,Fize,&Marlot,1996;VanRullen &Thorpe,2001a );a consider-able amount of information about which stimulus was shown is available in the spike counts of single inferior tem-poral cortex neurons in 20ms (Tovee &Rolls,1995);and we have shown in a backward masking paradigm that neu-rons in the inferior temporal cortex fire for only approxi-mately 30ms to a visual stimulus presented for 16ms which can be identified above chance (Rolls,2003;Rolls &Tovee,1994;Rolls,Tovee,&Panzeri,1999;Rolls,Tovee,Purcell,Stewart,&Azzopardi,1994).Delorme and Thorpe (2001)have suggested that just one spike from each neuron is sufficient,and indeed it has been suggested0042-6989/$-see front matter Ó2006Elsevier Ltd.All rights reserved.doi:10.1016/j.visres.2006.07.026*Corresponding author.Fax:+441865310447.E-mail address:Edmund.Rolls@ (E.T.Rolls).1Present address:Universidad de Malaga,Depto.de Lenguajes y Cs.de la Computacion,Campus de Teatinos S/N,29071Malaga,Spain./locate/visresVision Research 46(2006)4193–4205that the order of thefirst spike in different neurons may be part of the code(Thorpe,Delorme,&Van Rullen,2001; VanRullen,Guyonneau,&Thorpe,2005;VanRullen& Thorpe,2001b).An alternative view is that the number of spikes in afixed time window over which a postsynaptic neuron could integrate information is more realistic,and this time might be in the order of20ms for a single receiv-ing neuron,or much longer if the receiving neurons are connected by recurrent collateral associative synapses and so can integrate information over time(Deco&Rolls, 2006;Rolls&Deco,2002;Rolls et al.,1999;Tovee&Rolls, 1995;Tovee,Rolls,Treves,&Bellis,1993).Although the number of spikes in a short time window of e.g.20ms is likely to be0,1or2,the information available may be more than that from thefirst spike alone,and we examine this by actually measuring neuronal activity,and then applying quantitative information theoretic methods to measure the information transmitted by single spikes,and within short time windows.Moreover,we perform this analysis both for single neurons,and for populations of neurons,and measure whether more information is avail-able if the order of the spike arrival times from each neuron is taken into account in addition to whether any spike is present or not from each neuron.2.Methods2.1.Recording techniquesThe responses of single neurons in the temporal cortical visual areas were measured to a set of20visual stimuli in a rhesus macaque performing a visualfixation task using experimental procedures similar except as described below to those described in detail previously(Rolls,Treves,& Tovee,1997a).The stimuli included S=20grayscale images of objects (7),faces(8),natural scenes(3),and geometrical stimuli(2)of the type which produce differential responses from inferior temporal cortex neu-rons,and examples of which have been illustrated previously(Rolls& Tovee,1995).The resolution of these images was256wide by256high with256grey levels.The neurons were selected where possible to show responses that differed between the different stimuli(as shown by a one-way ANOVA).Usually20trials for each stimulus were available.The set of stimuli were shown once in random order,then a second time in a new random sequence,etc.Populations of2–9neurons were record-ed simultaneously using2–4independently movable single neuron epoxy-insulated tungsten electrodes with uninsulated tip diameters of less than10microns(FHC Inc.,USA)using an Alpha-Omega(Israel) recording system.Typically,we were able to move the microelectrodes until2–4of the simultaneously recorded neurons responded differen-tially to the set of stimuli used.This differentialfiring of the2–4neu-rons was evident in thefiring rates to the different stimuli.For the other typically2–4neurons being simultaneously recorded in the same brain region,there were no differentialfiring rates to the stimuli,but these were also included in some of the analyses as described below. The recordings were made as part of the experimental design in one rhesus macaque,Macaca mulatta,so that an analysis could be per-formed of non-simultaneously recorded neurons in which the informa-tion from all the recordings made from different neurons in different sessions in the same animal could be analysed as described by Rolls et al.(1997a).The microelectrodes were stereotaxically guided,and the location of the microelectrodes was reconstructed on each track using X-rays and subsequent histological reconstruction using microle-sions made on selected tracks as described by Feigenbaum and Rolls (1991).The recording system(Neuralynx Inc.,USA)filtered and amplified the signal and stored spike waveforms which were later sort-ed to ensure that the spike waveforms from each neuron in the small number of cases when there were more than two spikes on one micro-electrode were clearly separated into different waveform clusters using the Datawave(CO,USA)Discovery software.The neurophysiological methods used here have been described in detail by Booth and Rolls (1998).All procedures,including preparative and subsequent ones, were carried out in accordance with the NIH Guide for the Care and Use of Laboratory Animals,the guidelines of The Society for Neuroscience,and were licensed under the UK Animals(Scientific Pro-cedures)Act,1986.Eye position was measured to an accuracy of0.5degrees with the search coil technique(Judge,Richmond,&Chu,1980),and steady fixation of a position on the monitor screen was ensured by use of a(blink version of a)visualfixation task.The timing of the task is described below. The stimuli were static visual stimuli presented at the centre of the video monitor placed at a distance of53cm from the eyes.A full-size face image typically subtended21degrees in the visualfield.Thefixation spot posi-tion was at the centre of the screen.The monitor was viewed binocularly, with the whole screen visible to both eyes.2.2.Visualfixation taskEach trial started atÀ500ms(with respect to the onset of the test image)with a500ms warning tone to allowfixation of thefixation point, which appeared at the same time.AtÀ100ms thefixation spot was blinked offso that there was no stimulus on the screen in the100ms period immediately preceding the test image.The screen in this period,and at all other times including the inter-stimulus interval,was set at the mean lumi-nance of the test images.At0ms,the tone was switched offand the test image was switched on for500ms.At the termination of the test stimulus thefixation spot reappeared,and then after a random interval in the range 150–3350ms it dimmed,to indicate that licking responses to a tube in front of the mouth would result in the delivery of fruit juice.The dimming period was500ms,and after this,thefixation spot was switched off,and reward availability terminated500ms later.(A diagram of the timing of this task is provided by(Tovee&Rolls,1995;Tovee,Rolls,&Azzopardi, 1994).)The monkey was required tofixate thefixation spot in that if he licked at any time other than when the spot was dimmed,saline instead of fruit juice was delivered from the tube;in that the dimming was by so little that it could only be detected if the monkeyfixated the spot; and in that if the eyes moved by more than1degree from time0until the start of the dimming period,then the trial was aborted.(When a trial aborted,a high frequency tone sounded for0.5s,no reinforcement was available for that trial,and the inter-trial interval was lengthened from8 to11s.)2.3.Measuring the information from many recorded neurons using a decoding procedureEstimates of how much information was available from a population of neurons in afixed time window about which stimulus was shown were calculated using the decoding method described by Rolls et al.(1997a)to analyse the information.This method measures the amount of informa-tion available from the number of spikes from each neuron assuming non-simultaneous recording so that it can be applied to the data accumu-lated from different cells recorded in the same macaque over different days. The method can be used for very large numbers of cells,and when there are many spikes in a time window.The method uses a decoding procedure in which on each trial the probability that each stimulus(called s0)was shown is estimated from the vector of neuronal responses.This estimate is made by comparing the vector of neuronal responses on that trial to the average response vectors to each stimulus.Then,knowing the actual stimulus shown on that trial,the mutual information h I p i between the esti-mated stimulus s0and the real stimulus s over the set of stimuli S can be calculated as:4194 E.T.Rolls et al./Vision Research46(2006)4193–4205I p ¼X S s ¼1X S s 0¼1P ðs ;s 0Þlog 2P ðs ;s 0ÞP ðs ÞP ðs 0Þ:The decoding procedure used for the results presented here is Bayesian probability estimate (PE)decoding using a Gaussian fit,as described by Rolls et al.(1997a),Rolls and Treves (1998),Rolls and Deco (2002)and Franco,Rolls,Aggelopoulos,and Treves (2004),and includes cross-vali-dation procedure and bias correction procedures as described in detail by Rolls et al.(1997a)and Franco et al.(2004).The same analysis pro-gram was used for the calculation of the information from single cells.Further details of the decoding procedures are as follows (see also Rolls et al.(1997a)and Franco et al.(2004)).The full probability table estimator (PE)algorithm uses a Bayesian approach to extract P (s 0j r )for every single trial from an estimate of the probability P (r j s 0)of a stimu-lus-response pair made from all the other trials (as shown in Bayes’rule shown in Eq.(1))in a cross-validation procedure described by Rolls et al.(1997a)P ðs 0j r Þ¼P ðr j s 0ÞP ðs 0ÞP ðr Þ;ð1Þwhere P (r)(the probability of the vector containing the firing rate of each neuron,where each element of the vector is the firing rate of one neuron)is obtained as:P ðr Þ¼Xs 0P ðr j s 0ÞP ðs 0Þ:ð2ÞThis requires knowledge of the response probabilities P (r j s 0)which can be estimated for this purpose from P (r ,s 0),which is equal toP ðs 0ÞQc P ðr c j s 0Þ,where r c is the firing rate of cell c .We note that P ðr c j s 0Þis derived from the responses of cell c from all of the trials except for the current trial for which the probability estimate is being made.The probabilities P ðr c j s 0Þare fitted with a Gaussian (or Poisson)distribution whose amplitude at r c gives P ðr c j s 0Þ.By summing over different test trial responses to the same stimulus s ,we can extract the probability that by presenting stimulus s the neuronal response is interpreted as having been elicited by stimulus s 0,P ðs 0j s Þ¼Xr 2testP ðs 0j r ÞP ðr j s Þð3ÞAfter the decoding procedure,the estimated relative probabilities (normal-ized to 1)were averaged over all ‘test’trials for all stimuli,to generate a(regularized)table P R N ðs j s 0Þdescribing the relative probability of each pair of actual stimulus s and posited stimulus s 0(computed with N trials).From this probability table the mutual information measure I p was calculated as described above.Because the probability tables from which the information is calculated may be unregularized with a small number of trials,a bias correction pro-cedure to correct for the undersampling is applied,as described in detail by Rolls et al.(1997a)and Panzeri and Treves (1996).In practice,the bias correction that is needed with information estimates using the decoding procedures described here and by Rolls et al.(1997a,1997b)is small,typ-ically less than 10%of the uncorrected estimate of the information,pro-vided that the number of trials for each stimulus is in the order of the number of stimuli.We also note that the distortion in the information esti-mate from the full probability table needs less bias correction than that from the predicted stimulus table (i.e.maximum likelihood)method,as the former is more regularized because every trial makes some contribu-tion through much of the probability table (see Rolls et al.(1997a)).2.4.Measuring the information in the order of spike arrival timesThe decoding method as just described measures how much informa-tion is present in the first spike,or in the number of spikes in a given time window.It is possible that there is in addition extra information in the order in which the first spike arrives from different neurons about which stimulus is shown.Although normally the order of spike arrival would be expected to depend on the average firing rate of neurons (with for example a neuron with a high average firing rate to a particular stimuluslikely to produce an early spike),in principle the order information could be distinct from the rate information (with for example a neuron with a low firing mean firing rate to a stimulus nevertheless producing an early spike to that stimulus),thus providing additional information independent from the mean response about which stimulus was shown.We were able to test for this type of order information by ordering the first spike arrival times for each neuron on each trial to each stimulus,and measuring the information when it was this rank ordering that was being decoded by the dot product algorithm described below (see also Franco et al.,2004;Rolls et al.,1997a ).To measure the information contained in the relative order of the spikes from different neurons,for each recorded trial we constructed a vector with a number of components equal to the number of simultaneously recorded neurons.Each component contained the rank order of the first spike of that neuron relative to the other neu-rons.(The component contained one if the first spike from that neuron was the first among all neurons to arrive,2if it was the second,etc.If there was no spike from that neuron on that trial,the component was set to one greater than the number of spikes that had arrived.)To measure the infor-mation contained in the order of the arrival of the spikes of the different neurons,dot product decoding then compares each trial with the average of the trials to each stimulus,to determine how close the current trial is to the data obtained to each stimulus.(The average vector for each stimulus is calculated excluding the data for the current trial in what is a cross-val-idation procedure (Franco et al.,2004;Rolls et al.,1997a ).)The dot prod-uct (DP)algorithm computes the normalized dot product between the current firing vector r on a ‘test’(i.e.the current)trial and each of the mean firing rate response vectors in the ‘training’trials for each stimulus s 0in the cross-validation procedure.(The normalized dot product is the dot or inner product of two vectors divided by the product of the length of each vector.The length of each vector is the square root of the sum of the squares.)Thus,what is computed are the cosines of the angles of the test vector of cell rates with,in turn for each stimulus,the mean response vector to that stimulus.The highest dot product indicates the most likely stimulus that was presented,and this is taken as the best guess or (the predicted stimulus S p )for the probability table P (S ,S p )with max-imum likelihood information measurement (Rolls et al.,1997a ).We veri-fied that using this rank ordering in this information measurement procedure was effective by simulating spike trains with the same mean fir-ing rate of each neuron to any stimulus,and then introducing rank order-ing so that the order of spikes from the different neurons was different for each stimulus.The control condition was to take the same number of spikes in the time window in which the information was being measured,but to allocate the time of any spikes in that window to random times (with a uniform probability distribution),thus preserving any information present in the number of spikes in the time window (i.e.the ‘‘rate’’infor-mation),but removing any information present in the relative time at which spikes arrived from different neurons for each stimulus.In both cases,just the first spike in the time window was used in the information analysis.The reason for using just the first spike is that the order informa-tion becomes arbitrary after the first spike from each neuron,for it is not assumed that the order encoding in the brain acts like a counting processor that is able to take into account for example that the second spike from neuron x has to follow the first spike from neuron y .Indeed,when sug-gesting that order information was encoded,Delorme and Thorpe (2001)used just one spike per neuron.3.ResultsFrom 62cells recorded simultaneously in groups of 2–4neurons in 31experiments,we performed the analyses on 21neurons that had significant differences in their firing rate to the set of 20stimuli,as shown by ANOVA (p <.001).There were typically 20trials of data available for each stimulus for each neuron.Examples of peristimu-lus time rastergrams and histograms for a neuron are illus-trated in Fig.1.Most of the neurons had their bestE.T.Rolls et al./Vision Research 46(2006)4193–42054195responses to objects,and such neurons sometimes respond-ed to one or two faces in the set.We began by measuring the information in the first spike of neurons about which stimulus was present,and compared this with the information present in time win-dows starting at the same time but extending for longer durations.We did this for both single cells,and for multi-ple simultaneously recorded cells.This addresses the issue of how much information is available in the first spike com-pared to longer spike trains.To further analyze the speed and nature of cortical processing,we measured the infor-mation available in short fixed time windows (of 20and 50ms),and compared this to the information available from the first spike.To measure whether the order of the arrival times of the spikes of different neurons is part of the code,as has been hypothesized (Thorpe et al.,2001),we measured the information available when the order was retained,and compared it to the information present when the order of the spike times of each neuron for each stimulus was made random as described in Section 2.The cumulative single cell information about which of the 20stimuli was shown from all spikes and from the first spike starting at 100ms after stimulus onset is shown in Fig.2a.One hundred milliseconds is just longer than the shortest response latency of the neurons from which recordings were made,so starting the measure at this time provides the best chance for the single spike measurement to catch a spike that is related to the stimulus.The means and standard errors across the 21different neurons are shown.The cumulated information from the total number of spikes is larger than that from the first spike only,and this is evident and significant within 50ms of the start of the time epoch.In calculating the information from the first spike,just the first spike in the analysis window start-ing in this case at 100ms after stimulus onset was used.The brain does not know when to start looking for the first spike to a stimulus,so we show a re-analysis starting at the time at which the stimulus appeared in Fig.2b.As expected,the information measure starts to increase at approximately 100ms.The information measures do not in general reach such high values as when the analysis starts at 100ms,and the reason for this is that any spikes due to spontaneous neuronal firing in the period 0–100ms after stimulus presentation are effectively noise in the system.This effect is particularly a problem for the single spike measurement condition,for the system just utilizes the first spike after time 0,and this might well be a spontaneous activity-related spike with no information at all about the stimulus.This underlines the fact that the single spike measure does rather make the assumption that one knows when to start measuring the effects of the stimulus,if much information about the stimulus is to be extracted.Because any one neuron receiving information from the population being analyzed has multiple inputs,we show in Fig.3the cumulative information that would be available from multiple cells (21)about which of the 20stimuli was4196E.T.Rolls et al./Vision Research 46(2006)4193–4205shown,taking both thefirst spike after100ms,and the total number of spikes after100ms from each neuron (Fig.3a).The cumulative information even from multiple cells is much greater when all the spikes rather than just thefirst spike are used.The same conclusion is reached if the measurement starts at0ms with respect to the stimulus onset,as shown in Fig.3b.As in Fig.2,the information particularly with thefirst spike is less when starting at 0ms than at100ms,because in the0ms condition thefirst spike may not reflect anyfiring produced by the stimulus, as it may just be spontaneous activity.An attractor network might be able to integrate the information arriving over a long time period of several hundred ms,and might produce the advantage shown in Fig.3for the whole spike train compared to thefirst spike only.However a single layer pattern association network might only be able to integrate the information over the time constants of its synapses and cell membrane,which might be in the order of15–30ms(Panzeri,Rolls,Batta-glia,&Lavis,2001;Rolls&Deco,2002).In a hierarchical processing system such as the visual cortical areas,there may only be a short time during which each stage may decode the information from the preceding stage,and then pass on information sufficient to support recognition to the next stage(Rolls&Deco,2002).We therefore analyzed the information that would be available in short epochs from multiple inputs to a neuron,and show the multiple cell information for the population of21neurons in Fig.4a (for a20ms epoch)and in Fig.4b(for a50ms epoch). The epochs started at0ms with respect to stimulus onset. We see in this case that thefirst spike information,because it is being made available from many different neurons(in this case21selective neurons discriminating between the stimuli each with p<.001in the ANOVA),fares better rel-ative to the information from all the spikes in these short epochs,but is still less than the information from all the spikes,particularly in the50ms epoch.In particular,for the epoch starting100ms after stimulus onset in Fig.4a the information in the20ms epoch is0.37bits,and from thefirst spike is0.24bits.Correspondingly,for the 50ms epoch,the values in the epoch starting at100ms post stimulus are0.66bits for the50ms epoch,and0.40bits for thefirst spike.Thus,with a population of neurons,having just one spike from each can allow considerable informa-tion to be read if only a limited period(of e.g.20or 50ms)is available for the readout,though even in these cases,more information was available if all the spikes in the short window are considered(Fig.4a and b).To show how the information increases with the num-ber of neurons in the ensemble in these short epochs,we show in Fig.4c the information from different numbers of neurons for a20ms epoch starting at time=100ms with respect to stimulus onset,for both thefirst spike condi-tion and the condition with all the spikes in the20ms window.The linear increase in the information in both cases indicates that the neurons provide independent information,which could be because there is no redun-dancy or synergy,or because these cancel(Rolls,Agge-lopoulos,Franco,&Treves,2004;Rolls,Franco, Aggelopoulos,&Reece,2003b).It is also clear from Fig.4c that even with the population of neurons,and with just a short time epoch of20ms,more information is available from the population if all the spikes in 20ms are considered,and not just thefirst spike.The 20ms epoch analyzed for Fig.4c is for the post-stimulus time period of100–120ms.(Given this,the values for21 neurons in Fig.4c correspond to the values shown for time100–120ms in Fig.4a.)Thus,even with a population of neurons selected to be tuned to the stimulus set,at21 neurons the information available from thefirst spike only has not reached that obtainable from all the spikes in the20ms epoch.(Thefirst spike information measures the information obtained by the presence or absence of a spike in the given time window.)E.T.Rolls et al./Vision Research46(2006)4193–42054197To elucidate the spike train properties that give rise to the information just described,we show in Fig.5a the probability distribution of the number of spikes on each trial from each neuron for the most effective stimulus for each neuron,for a20ms epoch starting at100ms after stimulus onset.Because the data in Fig.5a is com-puted across all21neurons,we note that some individual neurons will have relatively higher probabilities of2or more spikes on each trial in this time window.For com-parison,we show in Fig.5b the probability distribution of the number of spikes on each trial from each neuron across all stimuli for each neuron,for the same20ms epoch.To assess whether there is information that is specifically related to the order in which the spikes arrive from the dif-ferent neurons,we computed for every trial the order across the different simultaneously recorded neurons in which thefirst spike arrived to each stimulus,and used this in the information theoretic analysis described in Section2. The control condition was to randomly allocate the order4198 E.T.Rolls et al./Vision Research46(2006)4193–4205。
LTF-C – Neural Network for Solving Classification Problems
Vectors X (i) can be treated as points in n-dimensional space X (we can identify vectors with points, so for the simplicity of the notation these terms will be used interchangeably). Close neighborhood of the point X (i) should belong to the same class as X (i) , therefore the space X can be divided into finite number of decision regions – areas of the same value of classification. The problem resolves then to the task of modeling complex figures in n-dimensional space. One of possible solutions of such a task is to model interiors of these regions – by filling them as tight as possible with figures of versatile shapes. This idea lays in the basis of LTF-C. The network is composed of two layers of neurons. The first one retains the information about figures filling the decision regions. Each figure is represented by a neuron (forms its reception field): the neuron weights define the position of the figure centre, and the radii – its size. The neuron output, belonging to the range of [0, 1], says how “much” the presented pattern lies in the interior of the figure. Formally, the response yi of the i-th neuron on the pattern X is given as: n 2 w − x ij j , yi = f (1) r ij j =1 where Wi = [wi1 , wi2 , . . . , win ] – weights of the i-th neuron, Ri = [ri1 , ri2 , . . . , rin ] – radii of the i-th neuron, f – an output function. With such a definition reception fields have the shapes of hyperellipses with axes parallel to the axes of the co-ordinate system. Reception fields have to fill some – the most often bounded – region, so they have to be bounded figures themselves. Hence the output function should satisfy:
Exploring the Frontiers of Neuroscience
Exploring the Frontiers of Neuroscience Exploring the Frontiers of Neuroscience Neuroscience is a field that has captivated scientists, researchers, and even the general public for decades. It delves into the intricate workings of the brain and the nervous system, seeking to unravel the mysteries of human cognition, behavior, and consciousness. As we push the boundaries of our understanding, new frontiers in neuroscience emerge,offering exciting possibilities and raising thought-provoking questions. One ofthe frontiers in neuroscience is the study of neuroplasticity, the brain's ability to reorganize and adapt. Traditionally, it was believed that the brain was a fixed and unchanging organ, but recent research has shown that it is far more malleable than previously thought. This discovery has profound implications for thetreatment of neurological disorders and the potential for enhancing cognitive abilities. For example, neuroplasticity has been harnessed to developrehabilitation techniques for individuals who have suffered from strokes or traumatic brain injuries, allowing them to regain lost functions or learn new ones. Another frontier in neuroscience is the exploration of the neural mechanisms underlying consciousness. Consciousness is a fundamental aspect of human experience, yet its origins and nature remain elusive. Advances in brain imaging techniques, such as functional magnetic resonance imaging (fMRI), have allowed researchers to observe the brain in action and correlate neural activity with conscious experiences. This has led to the identification of specific brainregions and networks that are involved in various aspects of consciousness, suchas self-awareness, attention, and perception. However, the question of how subjective experiences arise from neural activity is still a matter of intense debate and speculation. The study of the brain's connectome is also an exciting frontier in neuroscience. The connectome refers to the complete map of neural connections in the brain, akin to a wiring diagram. Mapping the connectome is a monumental task, as the human brain contains billions of neurons and trillions of connections. However, advancements in imaging technology and computational methods have made it possible to start unraveling this complex network. Understanding the connectome could provide insights into how information is processed andtransmitted in the brain, leading to a deeper understanding of brain function anddysfunction. Neuroscience also intersects with other fields, such as artificial intelligence (AI) and robotics, opening up new frontiers for exploration. By studying the brain, researchers can gain inspiration for developing more efficient algorithms and intelligent systems. Conversely, AI and robotics can serve as powerful tools for studying the brain, allowing researchers to simulate and manipulate neural networks in ways that would be impossible in living organisms. This interdisciplinary approach has the potential to revolutionize both neuroscience and AI, leading to advancements in fields such as brain-computer interfaces, neuroprosthetics, and cognitive enhancement. While these frontiers in neuroscience offer immense potential, they also raise ethical and philosophical questions. For example, as we gain a deeper understanding of the brain and its mechanisms, we may be confronted with the ability to manipulate and enhance human cognition and behavior. This raises concerns about the potential misuse of such technologies, as well as the implications for personal identity and autonomy. Additionally, the study of consciousness raises profound philosophical questions about the nature of the mind and its relationship to the physical brain. In conclusion, exploring the frontiers of neuroscience is an exciting and challenging endeavor. From understanding neuroplasticity to unraveling the mysteries of consciousness and mapping the brain's connectome, neuroscience offers a wealth of opportunities for discovery and innovation. By embracing an interdisciplinary approach and addressing the ethical and philosophical implications, we can navigate these frontiers with care and harness the power of neuroscience for the betterment of humanity.。
2021-2022学年南京市溧水区九年级上英语期末考试题.docx
2021〜2022学年度第一学期期末调研九年级英语注意事项:1.本试卷共10页。
全卷满分120分。
考试时间120分钟。
试题包含选择题和非选择题。
考生答题全部在答题卡上,答在本试卷上无效。
2.选择题必须用2B铅笔将答题卡上对应的答案标号涂黑。
答非选择题必须用0.5毫米黑色墨水签字笔卸载答题卡的指定位置,在其它位置答题一律无效。
选择题(共60分)第一部分:听对话,回答问题。
本部分共10道小题,每小题你将听到一段对话,每段对话听两遍。
在听每段对话前,你将有5秒钟的时间阅读题目,听完后,你仍有5秒钟的作答时间。
1.What does the man use to make beautiful music?2.Who was found dead in the horror film?4. Which programme is Tom watching?7. What can we know from the conversation?B.B.3. Where are the two speakers?5.What does the man want to do?A. Go dancing.6.What does the murderer look like?A. He has long hair.B. Listen to music.B. He has short hair.C. Play music.C. He is tall.A.The researchers measured the brain activity of 18 adults who were asked to look at an image (图像)on a screen for a certain length of time. They were then asked to guess how long they'd been looking at the image. They tended to (倾向于)guess incorrectly when they were asked to stare at the image for very long or very short periods of time.After looking at their brain activity, the researchers found neurons that send messages in response to certain amounts of time. When they receive repetitive stimulation (重复的刺激)(such as staring at a screen), these neurons little by little get "tired" and don't work properly. It doesn't matter whether this amount of time is long or short, as long as the stimulation is repetitive. However, other neurons still work normally, causing an imbalance in the way that we experience time.When you stand in line or do some other repetitive tasks, such as math homework, your time-sensitive neurons get tired, which may lead you to feel like time is going slowly. But when you're doing something more fast-paced, such as playing soccer, you feel like time is going faster.Either way, when it comes to your experience of time, think about whether your brain has made a mistake.Common Illusion (错觉)of time:Time seems to go at different speeds in different 81 . For example, a term may already have come toan 82 before we realize it.B)根据短文内容及首字母提示,填写所缺单词,并将答案填写在答题卡标号为96-105的相应位置。
- 1、下载文档前请自行甄别文档内容的完整性,平台不提供额外的编辑、内容补充、找答案等附加服务。
- 2、"仅部分预览"的文档,不可在线预览部分如存在完整性等问题,可反馈申请退款(可完整预览的文档不适用该条件!)。
- 3、如文档侵犯您的权益,请联系客服反馈,我们会尽快为您处理(人工客服工作时间:9:00-18:30)。
Yamaguchi and Kimizuka (20) tested the "singularity" of umami by presenting human subjects with 21 taste stimuli including single and mixture solutions of MSG and sucrose, NaC1, tartaric acid and quinine sulphate. The subjects sorted the stimuli based on taste quality similarity. These scores were placed into a similarity matrix and analysed using multidimensional scaling procedures. The results revealed that, within a three-dimensional tetrahedron, the four prototypical stimuli were located at the vertices of a tetrahedron. The mixtures containing 2, 3 or 4 prototypical stimuli were located on the edges or surfaces of the tetrahedron. However, MSG was located outside of the tetrahedron, implying that the taste of umami is qualitatively different from the four prototypical stimuli used. These findings raise the question of whether umami taste operates through channels in the primate taste system which are separable from those for the "prototypical" tastes sweet, salt, bitter, and sour. [Although the concept of four prototypical tastes has been used by tradition, there is increasing discussion about the utility of the concept, and increasing evidence that the taste system is more diverse than this: (4)]. To investigate the coding of information about umami taste in the primate nervous Байду номын сангаасystem, the experiments described here were performed with neurons recorded in the taste cortex of monkeys. The umami stimulus used was MSG. We attempted to answer the following questions. 1) Do cells respond preferentially to umami? Particular attention was paid to whether cells responded differently to MSG and sodium
L E S L I E L. B A Y L I S 1 A N D E D M U N D T. R O L L S 2
Department of Experimental Psychology, University of Oxford South Parks Road, Oxford OX1 3UD, England
BAYLIS, L. L. AND E. T. ROLLS. Responses of neurons in the primate taste cortex to glutamate. PHYSIOL BEHAV 49(5) 973979, 1991,--In order to investigate the neural encoding of glutamate in the primate, recordings were made from 190 taste responsive neurons in the primary taste cortex and adjoining orbitofrontal cortex taste area in macaques. Single neurons were found that were tuned to respond best to glutamate (umami taste), just as other cells were found with best responses to glucose (sweet), sodium chloride (salty), HC1 (sour), and quinine HC1 (bitter). Across the population of neurons, the responsiveness to glutamate was poorly correlated with the responsiveness to NaC1, so that the representation of glutamate was clearly different from that of NaCI. Further, the representation of glutamate was shown to be approximately as different from each of the other four tastants as they are from each other, as shown by multidimensional scaling and cluster analysis. Moreover, it was found that glutamate is approximately as well represented in terms of mean evoked neural activity and the number of cells with best responses to it as the other four stimuli, glucose, NaC1, HCI and quinine. It is concluded that in primate taste cortical areas, glutamate, which produces umami taste in humans, is approximately as well represented as are the tastes produced by: glucose (sweet), NaCl (salty), HC1 (sour) and quinine HC1 (sour). Taste cortex Orbitofrontal cortex Insular cortex Glutamate Umami Primate
JAPANESE cooks have long used sea tangles to enhance the flavor of foods. However, it was not until 1908 that Ikeda discovered that it was the glutamate in sea tangles that was responsible for the flavor enhancing effect (3). Umami, meaning "deliciousness" in Japanese, was the name that Ikeda gave the distinctive flavor arising from the glutamate salts. Umami is a taste common to a diversity of food sources including fish, meats, mushrooms, cheese and some vegetables. Within these food sources, it is the synergistic combination of glutamates and 5'-nucleotides that creates the umami taste (19,20). Monosodium L-glutamate (MSG), guanosine 5'-monophosphate (GMP) and inosine 5'-monophosphate (IMP) are examples of umami stimuli and are now widely available as food additives. Umami does not act by enhancing the tastes of sweetness, saltiness, bitterness or sourness in foods, but instead may be a flavor in its own right, at least in humans. For example, Yamaguchi (20) found that the presence of MSG or IMP did not lower the thresholds for the prototypical tastes (produced by sucrose, NaC1, quinine sulphate and tartaric acid), suggesting that umami did not improve the detection sensitivity for the four basic taste qualities. Also, the detection thresholds for MSG were not lowered in the presence of the prototypical taste stimuli. This suggests that the receptor sites for umami substances are different from those for other prototypical stimuli (20). (A synergistic effect was found when IMP was added to MSG in that the detection threshold for MSG was dramatically lowered.) Further,