Perspectives on Cognitive Neuroscience 1988-3884

8/2/2019 Perspectives on Cognitive Neuroscience 1988-3884

1/6

Reprint Series4 November 1988,Volume 242, pp. 741-745 SCIENCE

Perspectives on Cognitive NeurosciencePATRICIAS. CHURCHLANDND TERRENCE. SEJNOWSKI

Copyright O 1988 by the American Association for the Advancement of Science


2/6

Perspectives on Cognitive Neuroscience

How is it that we can perceive, learn and be aware of theworld? The development of new techniques for studyinglarge-scale brain activity, together with insights fromcomputational modeling and a better understanding ofcognitive processes, have opened the door for collabora-tive research that could lead to major advances in ourunderstanding of ourselves.

N UROSCIENCE AND COGNITIVE SCIENCE SHARE THE GOALof trying to understand how th e mind-brain works. In thepast, discoveries at the neuronal level and explanations atthe cognitive level were so distant th at each often see med of merelyacademic significance to the o ther . Sym bol processing models basedon the digital compute r have been unpro mising as a means to bridgethe gap between neuroscience and cognitive science, because theydid not relate to wha t was known abo ut nervous systems at the levelof signal processing. However, there is now a gathering convictionamong scientists that the time is right for a fruitful convergence ofresearch from hitherto isolated fields. Th e research strategy develop-ing in cognitive neuroscience is neither exclusively from the topdown, nor exclusively from the bottom up. Rather, it is a coevolu-P. S. Churchland is in the Department of Philosophy at the University of CaliforniaSan Diego, La JoUa, CA 92093. T. J. Sejnowski is at The Salk Institute, La Jolla, CA92093, and in the Department of Biology, University of California at San Diego, LaJoUa, CA 92093.

tionary strategy, typified by interaction among research domainswhere research at one level provides constraints, corrections, mdinspiration for research at other levels (1).

LevelsThere are in circulation at least three different notions of the term"levels," as it is used to d escribe scientific research, each no tiocarving the landsca pe in a different way-levels of analysis, levels oorganization, a nd levels of proce ssing.Levels o arrnlysis concern the conceptual division of a phenom enoin terms of different classes of questions th at can be asked ab out it.framework articulated by Marr and Poggio (2) drew upon thconception of levels in compu ter science and identified three level(i) the com putation al level of abstract problem analysis, decompos

ing the task into its main constituents (for example, determinatioof the three-dimensional structure of a moving object from succesive views); (ii ) the level of the algorithm, specifjring a formprocedure to perform the task by providing the correct output forgiven input; and (iii) the level of physical impleme ntation. Marr (maintained that computational problems of the highest level coulbe analyzed independently of understanding the algorithm thaperforms the computation. Similarly, he thought the algorithmproblem of the second level was solvable independently of undestanding its physical implementation.Some investigators have used the doctrine of independence tconclude that neuroscience is irrelevant to unde rstanding cognition4 NOVEMBER 1988 ARTICLES 74


3/6

Fig. 1. Structural levels of organization in the nervous system. The spatialscale at which anatomical organizations can be identified varies over manyorders of magnitude. (Left) Drawin g by Vesalius (33) of the hu man brain,the spinal column, and the peripheral nerves. (Right) Schematic diagramsillustrate ( to p) a processing hierarchy of visual areas in mon key visual vortex(34); (center) a small network model for the synthesis of oriented receptivefields of simple cells in visual cortex (35); and (bottom) the structure of achemical synapse (36). Relatively little is known about the properties at thenetwork level in comparison with the detailed knowledge we have ofsynapses and the general organization of pathways in sensory and motorsystem.

However, the independence that Marr emphasized pertained only tothe formal properties of algorithms, not to how they might bediscovered (4) . Computational theory tells us that algorithms can berun on d ifferent machines and in th at sense, and tha t sense alone, thealgorithm is indepen dent of the implementation. Th e formal point isstraightforward: since an algorithm is formal, no specific physicalparameters (for example, vacuum tubes o r c a 2 + ) are part of thealgorithm. That said, it is important to see that the purely formalpoint cannot speak to the issue of how best to discover thealgorithm used by a given machine, nor how best to arrive at thene~~robiologicallydequate task analysis. Certainly it cannot tell usthat the discovery of the algorithms relevant to cognitive functionswill be independent of a detailed understanding of the nervoussystem. Moreover, different implementations display enormousdifferences in speed, size, efficiency, and elegance. The formalindependence of algorithm from architecture is something we canexploit to build other machines once we know how the brain works,but it is not a guide to discovery when we d o not yet know how thebrain works. Knowledge of brain architecture also can be theessential basis and invaluable catalyst for devising likely and power-ful algorithms-algorithms that migh t explain how in fact the braindoes its job.Levels of'oyanizntiorr. H ow d o the t hree levels of analysis ma p o nt othe nervous system? There is organized structure at different scales:molecules, synapses, neurons, networks, layers, maps, and systems( 5 ) Fig. 1).The range of structural organization implies, therefore,that there are many levels of implementation and that each has itscompanion task description. But if there are as many types of taskdescription as there are levels of structural organization, this diversi-ty will be reflected in a multiplicity of algorithms th at c haracterizehow the tasks are accomplished. This in turn means that the notionof the algorithmic level is as oversimplified as the notion of theimplementation level. Structure at every scale in the nervous sys-tem-molecules, synap ses, neuro ns, netw orks, layers, maps, andsystems (Fig. 1)-is separable conceptually but not detachable

physically. Psychological phenomena may be associated wvariety of levels. Some p erceptual states such as the "raw " paintoothache migh t be a low-level effect, whereas atten tion may deon a variety of mechanisms, some of which can be found at theof local neural networks and others at the level of larger nsystems that reside in many different locations in the brain.Levels oj~yrocessing. his concept could be described as folThe greater the distance from cells responding to sensory inpuhigher the degree of inform ation processing. Thus the level assis a function of synaptic distance from the periphery. O nmeasure, cells in the primary visual area of the neocortexrespond t o orien ted bars of light are at a higher level than cells ilateral geniculate nucleus ( LG N ), which in tu rn are at a higherthan retinal ganglion cells.Once the sensory information reaches the cerebral cortex itout through cortico-cortical projections into a multitude of pastreams of processing. In the primate visual system 24 visual have been identified (6).Many (perhaps all) forward projectionaccompanied by a back projection, and there are even mafeedback projections from primary visual cortex to the LGN. Gthese reciprocal projections, it migh t seem tha t the processing ldo not really form a hierarchy, but there is a way to orderinformation flow by exam ining the layer of cortex into which fproject. Forwa rd projections generally terminate in the middle lof cortex and feedback projections usually terminate in the uand lower layers (7). However, we do not yet understandfunction of these feedback pathways. If higher areas can affecflow of information through lower areas, then the concepsequential processing must be modified.Th e hierarchical organization typica l of earlier sensory areas iadhered to everywhere. O n the contrary, th e anatomy of ass ocareas .and prefrontal cortex suggests a m ore "dem ocratic" orgation, and processing appears to take place in webs of strointeracting networks (8).Decisions t o act a nd the execution of pand choices could be the outcome of a system with distribcontrol rather than a single control center. Coming to grips systems having distributed c ontrol will require bo th new experital techniques and new conceptual advances. Perhaps more appriate metaphors for this type of processing will emerge studying models of interacting networks of neurons.

Color Vision: A Case StudyAs an illustration of fruitful interactions between psychologyphysiology on a problem in perception, we have chosen seexamples from color vision. Similar examples can also be lbunthe areas of learning and addiction (9) and sensory-motor inttion (10, l l) . Newton's ingenious prism experiment demonstthat white light can be decomposed into a mixture of wavelenand recombined t o recover the white light. This physical descriof color, however, did not satisfy artists, who were well awarethe perception of color involved complex spatial and temeffects. As Goethe pointed out in Zur Farbenlehre, dark shaoften appear blue. The physical description of color andpsychological description of color perception are at two difflevels: The link between them is at the heart of the problerelating brain to co gnition . Three exam ples will be given t o illuhow such links are being made in color vision.The knowledge that mixtures of only three wavelengths ofare needed to match any color led Young t o propose in 1802that there are only three types of photoreceptors. Qu ite a difftheory of color vision was later proposed by Herin g, w ho suggthat color perception was based on a system of color opponents

742 SCIENCE, VOL


4/6

for yellow versus blue, one for red versus green, and a separatesystem for black versus white (13). Convincing experiments andimpressive arguments were marshaled by supporters of these tworival theories fo r nearly a c entury. T he debate w as finally settled byphysiological studies proving that both theories were right-indifferent parts of the brain. In the retina th ree different types ofcolor-sensitive photoreceptors were fou nd, as predicted by Youn g,and the genes for the three cone photopigments have been se-quenced (14 ). In the thalamus an d visual cortex there are neuro nsthat respond to Hering's color opponents (15). Evidently, even atthis early stage of visual processing the com plexity of the brain maylead to puzzles that can only be settled by know ing h ow the brain isconstructed (16).Recent progress in solving the problem of color constancy is asecond example of converging physiological and psychologicalresearch. Red apples look red under a wide range of illuminationeven tho ugh the physical wavelengths imping ing on the retina varydramatically from daylight t o interior lighting. Insights into colorconstancy have come from artists, w ho manipulate color contrasts inpaintings, psychophysicists, who have quantified simultaneous con-trast effects (17), an d theorists, wh o have modele d the m (18). Col orconstancy depend s on being able t o co mpu te the intrinsic reflectanceof a surface independently o f the incident light. Th e reflectance of apatch of surface can be approximately comp uted by comp aring theenergy in wavelength bands c oming from the patch o f surface to theaverage energy in these ban ds from neig hboring and distant regionsof the visual f ield. Th e sig nature of a color-sensitive n euron that wasperforming this computation would be a long-range suppressiveinfluence fro m regions o f the visual f ield outside th e conventionalreceptive field. Neurons with such color-selective surrounds havebeen reported in visual cortex area V 4 (19); the first nonclassical

Fig. 2. Schematic dia-gram of anatomical con-nections and responseselectivities of neuronsin early visual areas ofthe macaque monkey.Visual information fromthe retina is split intotwo parallel streams atthe level of the lateralgeniculate nucleus(LGN) , the parvocellu-lar and magnocellular di-visions. The parvocellu-lar stream projects totwo divisions of primaryvisual cortex (Vl): thecytochrome oxidase-richregions (Blob) and cyto-chrome oxidase-poor re-gions surrounding theblobs (Interblob). Themagnocelluiar streamprojects to layer 4B ofV1. These three divi-

lnferotemporal Parietal

sions of V1 project intocorresponding areas of V2: the "thin stripe," "interstripe," and "thickstripes" of cytochrome oxidase-rich and -poor regions in V2. These areas inturn project to visual areas V3, V 4, and MT (midd le temporal area, alsocalled V5). Heavy lines indicate robust primary connections, and thin linesindicate weaker, more variable connections. Dotted lines indicate connec-tions that require additional verification. Not all projections from these areasto o ther brain areas are represented. The neurons in each visual area respondpreferentially to particular properties of visual stimuli as indica ted by theicons: Prism, tuned or oppo nent w avelength selectivity; Angle, orientationselectivity; Spectacles, binocular disparity selectivity or strong binocularinteractions; Pointing h and, direction of motion selectivity. [Reprin ted withpermission from (6)]

surrounds were found for motion-selective cells in area M T (20). Ithese neurons are necessary for color constancy then their losshould result in impairments of color vision. Bilateral lesions ocertain extrastriate visual areas in m an d o prod uce achromatopsia-atotal loss of color perceptio n (21)-although this conditio n usually found with othe r deficits and the d amaged areas may no t bhomologous with area V 4 in monkeys.Th e third example of a link between brain and cognition comefrom research on how form, motion, and color information arprocessed in the visual system. If different parts of the system arspecialized for different tasks, for example, for mo tion or co lor, thethere should be conditions under which these specializations arrevealed. Suppose the "color system" is good at distinguishincolors, but not much else, and, in particular, is poor at determininshape, depth, an d motion, whereas t l ~ e shape sy s ten~" s nosensitive to color differences but to brightness differences. Whebou nda ries are marke d only by colo r differences-all differences ibrightness are experimentally removed-shape detection shou ld bimpaired. Psychophysical research has shown that this is indeed thcase. Th e perceived mot ion o f equiluminant co ntours is degrade(22); form cues such as shape-from-sh ading are difficult to interp re(23), and perceived d epth in rand om-do t stereograms collapses (24Physiological and anatomical research has begun to uncover possible ex planation for these pheno mena (25). The separate processing streams in cerebral cortex mentioned earlier carry visuainformation abo ut different properties of objects (6, 26). In particular the predominant pathway for color information diverges frothose carrying information on motion and depth (Fig. 2). T hseparation is not perfect, however, bu t equiluminant stimuli providphysiologists with a visual "scalpel" for tracking d ow n th e correlateof perceptual co herence in different visual areas.The lessons learned from color perception may have significancfor studyin g other cognitive domains. S o far as we kn ow only small fraction of the neu rons in t he visual system respon d in a wathat corresponds to o ur perceptual report o f color. Th e locations ithe brain where links between physiological states and perceptuastates can be foun d vary from the retina to deep in the visual systefor different aspects of color perception (27). New experimentatechniques will be needed to study these links when t he informatiois encoded in a large population of interacting neurons ( 10, 28).

Techniques and Research StrategiesColor vision is a problem tha t has been studied for hundreds oyears; we know much less about the biological basis of otheperceptual an d cognitive states. Fortunately, ne w techniques, sucas regional b lood flow analysis with positron emission to mograp h(PET) and magnetic resonance imaging ( M R I ) are becominavailable for noninvasively measuring b rain activity in hu man s. Witthese techniques the large-scale pattern o f wha t is happ ening w herand when in the brain can be determined ; later , as techniques wit

higher resolution are developed they can be focused o n th e relevanareas to ask how the processing is accomplished.A us eh l way to g et an overview of the assorted techniques is graph them with respect to temporal and spatial resolution. Thpermits us t o identify areas where there d o n ot yet exist techniquto get access to levels of organization at those spatio-temperaresolutions and t o com pare their strengths and weaknesses (Fig. 3For example, it is apparent that w e lack detailed information aboprocessing in neural networks within cortical layers and columnover a wide range of time scales, from milliseconds to ho urs. The realso a pressing need fo r experimental techn iques design ed to add rethe dendritic and synaptic level of investigation in cerebral cortex

4 NOVEMBER 1988 ARTICLES 7


5/6

Withou t these data it will not be possible to develop realistic modelsof information processing in cortical circuits.

Although we nced experimental data concerning the properties o fneurons and behavioral data abo ut psychological capacities, we alsoneed to find models that explain ho w patterns of activity in neuronsrepresent surfaces, optical flow, and objects; how networks developand learn, store, and retrieve information; and how networksacconlplish sensorinlotor and other types of integration. Ideally,modeling and experimental research will have a symbiotic relation-ship, such that each informs, corrects, and inspires the other.

Although many diverse kinds of things are presented as modelsfor some part of the nervous system, it is useful to distinguishbetween realistic models, which are genuinely and strongly predic-tive of some aspect of nervous system dynamics or anatomy, and

1 simplifping models, which though not so predictive, demonstratethat the nervous system could be governed by specific principles./ Connectionist network models (29),which are simplifying models,are typically motivated by cognitive phenomena and are governedprimarily by computational constraints, while honorin g very generalneurobiological constraints such as number of processing un its andtime required t o perform a task. Accordingly, they are moreproperly considered demons tration s of what could be possible andsometimes wha t is no t possible. Realistic models of actual neuralnetworks, by contrast, are primarily motivated by biological con-straints, such as the physiological and anatomical properties ofspecific cell types (30). Despite their different origins and sources ofdominant constraints, simplifying models and realistic neural mod-els are both based on the mathematics of nonlinear dynamical

Fig. 3. Schematic illustration of the ranges ofspatial and temporal resolution of various experi-mental techniques for studying the hnction of thebrain. The vertical axis represents the spatialextent of the technique, with the boundariesindicating the largest and smallest sizes of' theregion from which the technique can provideuseful information. Thus, single-unit recordingcan only provide information from a small regionof space, typically 10 to 50 pm on a side. Thehorizontal axis represcnts the minimum and maxi-mum time intervals over which information canbe collected with the technique. Thus, actionpotentials from a single neuron can be recordedwith millisecond accuracy over many hours.Patch-clamp recording allows the ionic currentsthrough single ionic channels to be measured.Optical and fluorescent dyes that can reveal mem-brane potential, ionic concentrations, and intra-cellular structure have been used with high resolu-tion in tissue culture, where it is possible to obtaina clear view of single cells (37, 38). However,recordings from the central nervous system are

\ limited in resolution by the optical properties ofnervous tissue and only about 0. -mm resolutionhas been achieved (39). Confocal microscopy is a, I recent development in light microscopy thatcould be used for improving the resolution of thetechnique for three-dimensional specimens (40).ERP (evoked response potential) and MEG

Brain 2

Ma p 1

Column o

Neuron -2Dendrite -3Synapse

-4

systems in high-dimensional spaces (31). The common conceptuand technical tools used in these models should provide linbetween two rich sources of experimental data, and consequentlconnectionist and neural models have the potential to coevoltoward an integrated, coherent account of information processingthe mind-brain.

The ultimate goal of a unified account does no t require that it besingle model that spans all the levels of organization. Instead tintegration will probably consist of a chain of models linkinadjacent levels. When one level is explained in te rms o f a lower levthis does not mean that the higher level theory is useless or th at thigh-level phenomena n o longer exist. O n the contrary, explanatiowill coexist at all levels, as they do in chemistry and physics, geneticand embryology.

ConclusionsIt would be convenient if we could understand the nature

cognition without understanding the nature of the brain itseUnfortunately, it is difficult if not impossible to theorize effectivon these matters in the absence of neurobiological constraints. Tprimary reason is that computational space is consummately vaand there are many conceivable solutions to the problem of howcognitive operation could be accomplished. Neurobiological daprovide essential constraints o n computational theories, and thconsequently provide an efficient means for narrowing the searspace. ~ ~ u a l l ~ i m ~ o r t a n t ,he dat a are also richly suggestive in hin

MEG + ERP

\ Light Microscopy

(magn&oencephaldgraphyjecord the average electrical and magnetic activi-. ty over large brain regions and are limited to events that take place overabout 1 s (41). The temporal resolution of PET (positron emissiontomography) depends on the lifetime of the isotope being used, whichranges from minutes to an hour. It may be possible to achieve a temporalresolution of seconds with 150o study fast changes in blood flow by usingtemporal binning of the gamma ray events (equivalent to the poststimulustime histogram for action potentials) (42). The 2-deoxyglucose (2-DG)technique has a time resolution of about 45 min and a spatial resolution of0.1mmwith large pieces of tissue and 1 pm with small pieces of tissue (43).The 2-DG technique can also be applied to humans with PET (44). Lesions

- 3 - 2 - 1 0 1 2 3 4 5 6 7Millisecond Second Minute Hour DayTime (log seconds)

allow the interruption of function to be studied immediately after ablationthe tissue and for a long period of time after the ablation (21, 4Microlesion techniques make selective modifications with substances suchibotenic acid, which destroys neurons but not fibers of passage, andamino-phosphonobutyric cid, which selectively and reversibly blocks a claof glutamate receptors (46). Video-enhanced light microscopy has openedwindow onto dynamical activity within neurons, such as the recent visualition of axonal transport of organelles on microtubules (37, 47). AU of tboundaries drawn show rough regions of the spatio-temporal plane whethese techniques have been used and are not meant to indicate fimdamenLimitations.SCIENCE, VOL. 2


6/6

concerning what m ight really be going o n and w hat computationalstrategies evolution mig ht have chanced upon . Moreover, it is by nomeans settled what e&tly are the functional categories at' thecognitive levels, and therories of lower level function may well becrucial to the discovery of the nature of higher level organization.Accordingly, despite the fact that the brain is experimentally de-manding, basic neurob iology is indispensable in the task of discover-ing the theories that explain how we perform such activities asseeing, thinking, and being aware.On the other h and, the possibility that cognition will be an openbook once we understand the details of each and every neuron andits development, connectivity, and response properties is likewisemisconceived. Even if we could simulate, synapse for synapse, ourentire nervous system, that accom plishment, by itself, would not bethe same as understanding how it works. The simulation might bejust as much of a mystery as the hn ct io n o f the brain currently is, forit may reveal nothing a bo utt he network and systems properties thathold th e key to cognitive effects. Even simulations of small networkmodels have capabilities that are difficult to understand (32).Genuine theorizing abou t the nature of neurocom putation is there-fore essential.Many major questions remain to be answered. Although someproblems in vision, learning, attention, a nd sensorim otor control areyielding, this will be harder to achieve for more co mplex psychologi-cal phenomena such as reasoning and language. onet the less, or&we understand some fundamental principles of brain function, wemay see how to reformulate the outsta nding problems and addressthem in ways that are impossible now to predict. Whatever theoutcome, the results are likely to surprise us.

RE F E RE N CE S A N D N O T E S1. P. S. Churchland, LVereurophi losophy: Toward a Utr i f ied Sc ience o jrhe Miud-Brai~r M ITPress, Cambridge, MA, 198 6); J. LeDoux and W. Hirst, M ~ r d nd Brain: Dia lo~qrres

i t1 Cognir ive Ner trosde trce (Cambridge Univ. Press, Cambridge, 1986 ); S. M.Kosslyn, Science 240, 162 1 (1988) .2. Th e original conception of levels of analysis can be found in D. Marr and T . Poggio[Netr rosd . Res . Program Btr11 . 15, 470 (1 977 )l. Although Marr (3 ) emphasized theimportance of the computational level, the notion of a hierarchy of levels grew o utofearlier work by W. Reichardt and T. Pogg io [Q. R e v . B i o p l r y s . 9 ,31 1 (1976) l onthe visual control of orientation in the fly. In a sense, the current view on theinteraction between levels is not so much a departure from the earlier views as areturn t o the practice that was previously established bv Reichardt. Pog gio, andeven by Marr himself, who published a series of papers on neural netwo rk modelsof the cerebellar cortex and cerebral cortex. See, for example, D. Marr, J. Plrys io l .( L o n d o n ) 202, 437 (1969); P r o c. R . S o t . L o tr do tr B 176, 161 (1970 ). The emphasison the computational level has nonetheless had an important influence on theproblems and issues that concern the current generation of neural and c onnection-ist models [T. J. Sejnowski, C. Koch, P. S. Churchlan d, Srietrce 241, 1299 (1988)l .3. D. Marr, V i s i o n (Freeman, San Francisco, 1982).4. P. M. Churchland, D i a l o p t e 21, 223 (1982) .5. F. H. C. Crick, S ci . A n . 241, 219 (September 1979 ); P. S. Churchland and T. J.Sejnowski, in Ne~tral o t rrrec t io t rs and Me~rta l o tnpr trar io t t , L. Nadel, Ed. (MI T Press,Cambridge, MA, 1988); G. M. Shepherd, Y a l e J. B i o l . M e d . 45, 58 4 (1972); P .Grobstein, B r ah r B e h a v . E v o l . 31. 34 (1988) .6. E. A. DeYoe and D. C. Van Essen, Trend s i \ in trosc i . 11, 219 (1988) .7. J. H. R. Maunsell and D . C. Van Essen, J . Nerrrosd . 3, 2563 (1983) .8. P. S. Goldman-Rakic, A t r t r u . R e v . N e t r r o s d . 11, 137 (1 988); V. B. Mountcastle,Treuds Nerrrosc i . 9, 505 (1986) .

9. T. H . Brown, P. Chapman, E . W. Kairiss, C. L. Keenan, Science 242, 724 (1988);G. F. Koob and F. E. Bloom, i b i d . , p. 715 ; L. R. Squire, Metnory and Brai t t (OxfordUniv. Press, Oxford, 1 987 ); E. R. Kandel et a/. , in Synapt ic Frorcr ion , G . M .Edelman, W. E. Gall, W. M. Cowan, Eds. (Wilev, New York, 1 987 ), pp. 471-518;M. Mishkin and T. Appenzeller, S t i . A t n . 256, 80 (June 1987) .10. S. P. Wise and R. Desimone, S c i a t v 242, 736 (1988) .11. S. G. Lisberger, ib id . , p. 728.12 . T. Younz. P h i l o s . T r a n s . R . S o c . L o ~ rd o r r92. 12 (18021.13. E . HeriG, Z u r Lehre vo tn Lic lr t si r tt r ( ~ e r l i n , ' 1 8 7 8 ) . '

J. Nathans, T. P. Piantanida, R . L. Eddy, T. B. Shows, D. S. Hogness, Sciorce203 (1986) .The possibility of a two-stage analysis for color vision was suggested as ear1881, but no progress was made until physiological techniques became avaifor testing the hypothesis. There are still some issues that have not yet been resolved by this theory.H . B. Barlow, Q.J. E x p . P s y c h i d . 37, 121 (1985) .D . Jameson and L. M. Huwich, J. O p r . So c. A n . 51, 46 (1961) .E. H. Land, P ro c. N a t l . A t a d . S c i. U . S . A . 83 , 3078 (19 86); A. Hurlbert anPoggio. Science 239, 482 (1988) .S. Zeki, Nerrroscisrce 9, 741 (1.983); R. Desimone, S. J. Schein, L. G . UngerleV i s i o ~ rR e s . 25, 441 (1985) .J. Allman, F. Miezin, E. McGuiness, 4111rrr.Rev. i\hrrosci. 8, 407 (1985) .A. R. Damasio, in Pri trc ip les q fBe11 avio rd .Verrr o lqy , M. M. Mesulam, Ed. (DPhiladelphia, 1985), pp. 259-288; 0 . Sacks, R. L. Wasserman, S. Zeki. RSiegel, S o c . N e w o s r i . A b s t r . 14, 1251 (1988) .V. S. Ramachandran and R. L. Gregory, Xatarlrre 275, 55 (1978)P. Cavanagh and Y. Leclerc, I n v e s t . O p h t h a l m o l . S u p p l . 26, 282 (1985) .C. LU and D . H . Fender, ib id . 11, 482 (1972) .M. S. Livingstone and D. H . Hubel, S c i e ~ t c e240, 740 (1988) .S . E k i , Nart t re 335, 311 (1988) .D. Y. Teller and E. N. Pugh, Jr., in Col or Vi .

Documents

Perspectives on Cognitive Neuroscience 1988-3884