Distributed Modular Architectures Linking Basal...

Feature Article

Distributed Modular ArchitecturesLinking Basal Ganglia, Cerebellum, andCerebral Cortex: Their Role in Planningand Controlling Action

The motor system includes structures distributed widely through theCNS. and in this feature article we present a scheme for how theymight cooperate in the control of action. Distributed modules, whichconstitute the basic building blocks of our model, include recurrentloops connecting distant brain structures, as well as local circuitrythatmodulates loop activity. We consider interconnections among thebasal ganglia, cerebellum, and cerebral cortex and the specializedproperties of certain cell types within each of those structures. namely. striatal spiny neurons, cerebellar Purkinje cells, and neocorticalpyramidal cells. In our model, striatal spiny neurons of the basal ganglia function in contextual pattern recognition under the training influence of reinforcement signals transmitted in dopamine fibers. Cerebellar Purkinje cells also function in pattern recognition, in theircase to select and execute actions through training supervised byclimbing fibers, which signal discoordination. Neocortical pyramidalcells perform collective computations learned through a localtrainingmechanism and also function as information stores for other modularoperations. We discuss how distributed modules might function in aparallel. cooperative manner to plan, modulate, and execute action.

Modularity has emerged as an important architectural principle of vertebrate nervous systems, one probably reflectingboth its evolutionary history and development (Rakic, 1988).The CNS has enlarged dramatically since vertebrate brainsfirst evolved, and replication of existing modules, subject tosubsequent phylogenetic and ontogenetic modification, represents a common and convenient mechanism for expansionand elaboration in biological systems. Columns in the neocortex (Mountcastle, 1978), parasagittal strips in the cerebellar cortex (Oscarsson, 1980; Voogd and Bigare, 1980), andstriosomes of the striatum (Graybiel, 1991) epitomize someof the many forms modularity assumes in the mammalianCNS. To that list one might add the barrels, blobs, bands, islands, pools, clusters, and aggregates of many kinds scatteredthroughout the vertebrate brain.

Neurobiologists typically consider modular organization inthe context of local circuits and circumscribed brain regions.However, principles of modularity may also apply to the organization of the neural networks that link physically separated regions of the CNS. This conjecture receives supportfrom neuroanatomical studies that demonstrate a high degreeof topographic specificity in the projection pathways linkingdifferent regions of the cerebral cortex, thalamus, basal ganglia, and cerebellum (e.g., Asanuma et al., 1983; Goldman-Rakic, 1984, 1988a,b; Ito, 1984; Alexander et al., 1986; Gibson etal., 1987; Shinoda and Kakei, 1989; Hoover and Strick, 1993).Motivated by these and other examples of anatomical specificity in projection pathways, we and others have been exploring an extension of the concept of modularity-frommodules restricted to a single neuronal structure to those thatmay be distributed over a number of distant, but functionallycooperative, structures (Goldman-Rakic, 1988a; Selemon andGoldrnan-Rakic, 1988; Houk, 1989; Houk et al., 1990; Eisenmanet al., 1991; Houk and Barto, 1992; Berthier et al., 1993; Houkand Wise, 1993; Wise and Houk, 1994). Specificity in these

James c. Houk! and Steven P.Wise 2

1 Department of Physiology, Northwestern UniversityMedical School, Chicago, Illinois 60611 and 2 Laboratory ofNeurophysiology, National Institute of Mental Health,Poolesville, Maryland 20837

long-distance connections may create modules with the interesting property of being anatomically distributed while being united in some emergent function. We term suchcooperatively interacting subunits distributed modules to reinforce the idea that they include disparately located neuronalstructures connected through their projection neurons.

The concept of distributed modules may assist understanding the cooperative function of the diverse "motor control"structures and the computational architectures that subserveadaptive action. In venturing to build on the work of manyothers who have considered the cooperative operations ofmotor control structures (e.g., Kornhiiber, 1971; Allen andTsukahara, 1974; Arbib, 1987; Kawato et al., 1987; Selemonand Goldman-Rakic, 1988; Grossberg and Kuperstein, 1989),we have proposed certain general information processingfunctions for three kinds of distributed modules: (1) corticalbasal ganglionic modules linking cerebral cortex, thalamus,and basal ganglia; (2) cortical-cerebellar modules linkingparts of the cortex, pons, thalamus, and cerebellum; and (3)cortical-cortical and their associated cortical-thalamic modules linking columns of cerebral cortex and thalamus throughreciprocal connections (Houk and Wise, 1993; Wise and Houk,1994). We begin our discussion by considering certain cellularspecializations in each of these three types of modules.

Cellular Specializations for Information ProcessingNeurons have impressive morphological and physiologicalspecializations (Llinas, 1988; Shepherd, 1990), which probablyreflect their distinct information processing capabilities. Although we cannot review this topic in detail, it will be important to highlight a few prominent cellular specializationsas they relate to the information processing modules discussed in this article.

Purkinje and Striatal Spiny CeUsThe classic theoretical work on pattern classification networks (Nilsson, 1965) and perceptrons (Rosenblatt, 1962;Minsky and Papert, 1969) employed threshold logic units.These artificial neurons had (1) high convergence ratios ofdiverse inputs onto each output neuron, (2) specialized training signals for the efficient adjustment of the large numberof weights associated with these inputs, and (3) sharp thresholds between on- and off-states of the output neurons. Thelatter property ensured clean "decision surfaces" for detectingthe presence or absence of the input patterns to which theunits were tuned. While pattern classifiers constructed fromsuch simple threshold logic units have definite limitations(Minsky and Papert, 1969), they nevertheless demonstrate remarkable recognition capabilities. Purkinje cells in the cerebellar cortex and striatal spiny neurons in the basal gangliahave several unique structural and functional properties thatappear to be analogous to threshold logic units.

First, like threshold logic units, Purkinje and striatal spinyneurons have high convergence ratios. This characteristic results from their extensive dendritic trees, numerous spinous

Cerebral Cortex Mar/Apr 1995;2:95-110; 1047-3211/95/$4.00

synapses, and a pattern of innervation favoring the convergence of diverse afferents onto individual neurons. Each Purkinje cell is contacted by approximately 200,000 differentparallel fibers (Ito, 1984), and each spiny neuron is contactedby about 10,000 different corticostriatal afferents (Wilson,1995). Those are the highest convergence ratios known inthe nervous system.

Second, Purkinje and spiny neurons receive specialized inputs that appear to transmit training information for guidingthe adjustment of synaptic weights. While many types of neuron exhibit synaptic plasticity, in most cases it appears to becontrolled by a Hebbian-like mechanism that depends mainlyon local presynaptic-postsynaptic correlation. As discussedlater, a Hebbian mechanism functions well for some types oflearning, such as the unsupervised learning problems confronted by sensory systems. By itself, however, it is inadequatefor training motor systems to interact with the environmentin useful ways. In these cases, training information that reports on the success or failure of interactions with the environment is also required, and learning is most efficient whenthis information is specific. In the cerebellar cortex, climbingfibers are highly specific and appear to transmit training information resembling a punishment signal reporting on failures in movement coordination (Ito, 1984; Houk and Barto,1992). In the basal ganglia, dopamine fibers appear to providetraining information that signals predictions of reinforcement(Wickens, 1990; Ljungberg et aI., 1991, 1992; Schultz et aI.,1993; Houk et aI., 1995). This input is topographically organized (Graybiel, 1991; Gerfen, 1992); however, the specificityis probably not as pronounced as that for climbing fibers.

Third, an unusually high density of voltage-dependent ionchannels render the electrical responses of Purkinje andspiny neurons exceptional. In particular, persistent calciumcurrent in Purkinje cell dendrites supports plateau potentialsthat have sharp thresholds for state transitions (Llinas andSugamori, 1980; Ekerot, 1984). This nonlinear dynamical feature is thought to convert dendrites into bistable elementsand the neurons as a whole into multistable input-outputunits (Houk et al., 1990; Yuen et aI., 1992). The discretethresholds for plateau potentials are analogous to the sharpdecision surfaces of threshold logic units. Similarly, the sharpdivision between "up" and "down" states exhibited by spinyneurons (Wilson, 1990) would provide sharp decision surfaces for these neurons in pattern recognition tasks.

Cortical Pyramidal CellsPyramidal neurons of the cerebral cortex also have severalimportant specializations in morphology and physiology, onesthat differ from those of Purkinje and spiny neurons. First,pyramidal neurons have combinations of ion channels thatpromote graded frequencies of firing in response to gradedstrengths of input (McCormick et aI., 1985), in contrast to thedual-state behavior characteristic of striatal spiny neurons andPurkinje cell dendrites. Processing units with graded outputsare clearly advantageous when the results of a neural computation need to be expressed over a range of magnitudes, ascontrasted with the "yes" or "no" values required of thresholdlogic units. Second, intra-axonal staining (Gilbert and Wiesel,1983; De Felipe et aI., 1986; Shinoda and Kakei, 1989) and other methods (e.g., Goldman and Nauta, 1977;Jones et aI., 1978;Goldman-Rakic, 1988a,b; Huntley and Jones, 1991) haveshown that individual fibers from the thalamus or cortex formclusters of synapses, which suggests that individual afferentscould make a variable number of synapses on a given pyramidal neuron. This would extend the range of input weights.Third, instead of specific training signals, pyramidal neuronsappear to receive several diffuse inputs that may provideglobal mechanisms for nonspecifically modulating synaptic

96 Distributed Modules in Motor Control' Houk and Wise

plasticity and learning (Bassant et aI., 1990; Waterhouse et al.,1990; Murphy et aI., 1993). At the cortical level learning appears to be guided mainly by local correlations between preand postsynaptic activity. Local learning rules are useful inunsupervised learning tasks, which can be effective in organizing sensory representations into topographic and cognitive(e.g., spatial) maps (Barlow, 1989; Reiter and Striker, 1989;Bear et al., 1990; Schlagger et aI., 1993) or in establishing automatic responses such as conditioned reflexes (Houk andBarto, 1992). For example, if a neuron already receives astrong, previously established connection (analogous to anunconditioned stimulus) capable of forcing the cell to respond in an appropriate manner, this strong input can thentrain other inputs by "example." In this manner, a neuron canselect from a set of weak inputs those pertinent to the taskat hand. Later, we suggest that this type of learning may occurin the frontal cortex where it could be guided by examplesetting, forcing inputs transmitted from the basal ganglia andcerebellum.

Cortical Basal-Ganglionic Modules

Context Recognition by Spiny NeuronsIn our model, the emergent functions of cortical-basal ganglionic modules stem from the capacity of their striatal spinyneurons for pattern recognition. In the cortical-basal ganglionic module illustrated in Figure lA, note the convergenceof input from several cortical columns (C for cortical columnsgenerally, F for a frontal column) onto a cluster of striatalspiny neurons (SP). This feature is well suited for contextualpattern recognition (Houk and Wise, 1993; Wise and Houk,1994). We postulate that, through the mediation of reinforcement training signals provided by the dopaminergic cells ofthe midbrain, this neuronal architecture learns to recognizeand register complex contextual patterns that are relevant tobehavior. This contextual information includes the state of theorganism, the desirability of an action, the actions planned inthe near future, the location of targets of action, and sensoryinputs suitable for either selecting or triggering motor programs.

Convergence onto Striatal Spiny CellsThe convergent input upon striatal spiny cells from diverseregions of the cerebral cortex is processed and fed back, viathe pallidum (P) and thalamus (T), to the frontal cortex (Fig.1A; Alexander et aI., 1986, 1990). The precise extent of cortic ostriatal convergence may vary substantially. Parthasarathyet al, (1992) have reported evidence that the terminal fieldsof projections from the frontal eye field (FEF) and the supplementary eye field (SEF) to the striatum overlap almost exactly. Similarly, Flaharty and Graybiel (1991) have shown thatparts of the somatosensory cortex and the primary motorcortex send converging inputs to striatal targets, potentiallyonto individual spiny cells. In contrast, Selemon and GoldmanRakic (1985) found that whereas prefrontal and posterior parietal inputs may project to the same general sector of thestriatum (Yeterian and Van Hoesen, 1978), they interdigitateat the local level rather than overlap. It is possible that bothinterdigitation and frank overlap can be found among corticostriatal projections, and that both provide for a diversity ofinputs to the striatal spiny cells that they contact.

Cortical-Basal Ganglionic Modules asContext DetectorsSome of the signals that cortical-basal ganglionic modulesmight support are illustrated in Figure lB. The presentscheme is based on Chevalier and Deniau (1990). We furthersuggest that the spiny neuron burst illustrated in trace SP

Acontext

registration

Figure 1. A, A cortical-basal ganglionicmodule, with an embedded cortical-thalamic module. C, cortical column composed ofcorticocortical, corticothalamic,and other projection neurons; F, a cortical column inthe frontal cortex; Sp' spinystriatal neuron; ST, subthalamic nucleusneuron; P, pallidothalamic neuron; T, thalamocortical neuron. B, Activation patterns of some of the elements in a cortical-basal ganglionic module. Traces SP,P, and T-F represent the discharge rateas a function of time for striatal spinyneurons, pallidothalamic cells, and frontal cortical-thalamic modules, respectively. For traces P and T-F, solid linesindicate signals transmitted through thedirect pathway in A and stippled linesindicate signals transmitted through theST side loop.

B

SP context detected

pallidal discharge patternsp

_nL_"__reflects the neuronal "recognition" of the context to whichthat spiny neuron is tuned. Spiny neurons inhibit pallidal neurons (cell P in Fig. lA), so their bursts are believed to producepauses in pallidothalamic activity (solid trace P in Fig. IB)(DeLong and Georgopoulos, 1981; Tremblay and Filion, 1989;Chevalier and Deniau, 1990). Pallidothalamic cells are also inhibitory, and one would expect their pause in discharge toinitiate a brief burst of thalamic discharge, due to the inhibitory rebound properties of these neurons (Wang et aI., 1991).This rebound burst might then initiate positive feedback inthe reciprocal cortical-thalamic loop shown in Figure lA(Houk and Wise, 1993), assuming that the corticothalamicprojection has a net excitatory influence (Ghosh et aI., 1994).Once initiated, positive feedback could be self-sustaining and

thus could continue after pallidal discharge resumed its tonicrate. Sustained cortical-thalamic loop feedback (solid trace TF in Fig. IB) would register that a relevant context recentlyoccurred, thus serving as a working memory (Fuster and Alexander, 1973; Goldman-Rakic and Friedman, 1991) of thecontextual information encoded by the striatal spiny neuron.This model bears important similarities to Hikosaka's (1989)views on dynamic interactions in the eye-movement controlsystem.

Strictly cortical mechanisms, in addition to thalamic disinhibition, undoubtedly contribute to the regulation of loop activity. For example, cortical inputs could initiate and/or helpsustain positive feedback in the cortical-thalamic modules.Consistent with this idea, cortical rather than pallidal afferents

Cerebral Cortex Mar/Apr 1995, V 5 N 2 97

have been implicated as causal factors in the premovementactivation of thalamic neurons that receive both inputs (Anderson and Turner, 1991). The pallidothalamic segment of acortical-basal ganglionic module would, on this view, be ideally situated to modulate discharge in cortical-thalamicmodules (Chevalier and Deniau, 1990), and would do so onthe basis of contexts recognized by the striatal spiny cells.

Some behavioral neurophysiology of the basal ganglia appears to support the proposal that it plays a role in contextrecognition and registration. Nearly all of the investigatorswho have studied striatal or pallidal modulation in awake,behaving monkeys have emphasized its context dependency(Alexander, 1987; Hikosaka et al., 1989b; Alexander andCrutcher, 1990; Schultz and Romo, 1992). Context-dependentneuronal activity, in this sense, indicates discharge modulations that do not obligatorily follow sensory events or accompany specific motor acts, but rather do so only under certaincircumstances. Context-dependent striatal modulation bothfollows sensory signals and accompanies movements (Rolls etal., 1983; Schultz and Romo, 1988, 1992; Hikosaka et al., 1989a;Romo and Schultz, 1992). Kimura (1992, p 212) concludedthat "most of the sensory responsive putamen neurons [81%]showed clear differences in the responses to identical stimuli... presented with[in] different behavioral contextjs},' Thisfinding includes the majority of spiny cells in the putamen(Kimura et al., 1990; see also Kimura, 1992), as well as thetonically active cells that are likely interneurons (Kimura etal., 1984; Apicella et al., 1991). Interestingly, this poststimulusactivity does not depend dramatically on movement parameters, including the direction of movement. In contrast to thisapparently sensory activity, which follows stimuli, many striatal modulations precede movements. This "movement-related" activity may reflect movement parameters (DeLong andGeorgopoulos, 1981), but it also depends upon the contextin which an action occurs (Kimura, 1990; Gardiner and Nelson, 1992). Again quoting Kimura (1990, p 1295), a fundamental "characteristic of ... movement-related cell activity[is] dependence on the conditions in which movement occurs. . . ." Brotchie et al. (1991) stress a similar opinion concerning the pallidal cells that receive striatal input. Contextdependency in striatal and pallidal activity accords with theidea that striatal spiny neurons act as pattern detectors.

Context Recognition and Mutual InhibitionOur notion of context recognition resembles Shallice's (1988)ideas concerning "contention scheduling." In that scheme,neuronal assemblies, activated by a pattern of sensory andother inputs, form mutually inhibitory networks that constitute a "winner-take-all" system. The module most activatedwill influence the action with which it is associated. Frith(1992) has previously suggested the basal ganglia as the siteof contention scheduling. This hypothesis accords well withthe elaborate system of spiny neuron GABAergic collaterals(Wilson and Groves, 1980; Groves, 1983; Aronin et al., 1986;Kawaguchi et al., 1990) postulated to form competitive networks within inhibitory domains (Alexander and Wickens,1993). Feedforward inhibition through local GABAergic interneurons might also contribute to these operations (pennartzand Kitai, 1991).

Through these inhibitory mechanisms, one can anticipatea high degree of mutual exclusion among cortical-basal ganglionic modules, which would result in a sparse coding ofsalient context. Thus, any given context would tend to berecognized by a relatively small number of modules, althoughthey might share many inputs. This concept has some important consequences when combined with the knowledge thatdifferent populations of striatal neurons may originate the "direct" and "indirect" pathways (Gerfen et al., 1990; Gerfen,


1992; Flaharty and Graybiel, 1993). The direct pathways inhibit the pallidothalamic output cell P in Figure lA, whereasthe indirect pathway relays inhibitory inputs from SP neuronsto a part of the globus pallidus (not shown) that in turn inhibits the excitatory ST input to P neurons (see Hazrati andParent, 1992). The net effect is thus disinhibitory, as indicatedby the open arrow in Figure lA. In our scheme, spiny neuronsof the direct pathway can be tuned to contexts for action,whereas nearby spiny neurons associated with the indirectpathway would respond to contexts for withholding action.As shown by the stippled traces in Figure IB, activity in theindirect pathway would enhance P discharge, rather than suppressing it, which could counteract context registration promoted by a direct pathway or, as illustrated in Figure IB,could negate a previously registered context. In accord withthis idea, some of the putamen neurons appear to be moreassociated with the withholding, rather than the execution,of movements (Hikosaka and Wurtz, 1983: Schultz and Romo,1992).

Function and Dysfunction ofCortical-BasalGanglionic ModulesOur suggestion that striatal spiny neurons function in contextrecognition appears to contrast with prevailing theoriesabout the role of the basal ganglia in motor control. Traditionally, the basal ganglia have been thought of as modulatorsof movement that release motor responses through a disinhibitory mechanism (Marsden, 1982; Hikosaka and Wurtz,1983; Albin et al., 1989; DeLong, 1990). By contrast, it has beenproposed that the role of the basal ganglia involves preventing maintained motor activities, including normal posture(Mink and Thach, 1991), from interfering with shifts in limbposition. The dual pathways for enabling and withholdingmovements, the direct and indirect pathways mentionedabove, may help to reconcile these apparently contradictoryviews. Since context is clearly important in determiningwhether to enable or suppress an action, both of these functions would fit well with the idea of context detection promoted here. Thus, suppression of postures (Mink and Thach,1991) can be viewed as a function of recognizing a contextfor withholding a rigid postural command and conveying thatrecognition to the thalamus.

Although we recognize the speculative nature of the assignment of a context recognition role to the basal ganglia,that idea accords with current models of its pathophysiology(Hallett, 1993). Ablating or inactivating the globus pallidusdoes not lead to paralysis. Movements can still be selectedand executed in an appropriate context (Horak and Anderson,1984; Mink and Thach, 1991). Not only can the motor systemstill select the correct action, but it does so without a dramatic increase in reaction time. This fact appears to contradict the view that the basal ganglia serve as context recognizers for action. If the system cannot recognize the contextfor action, one might ask, how can it act at all? However, thesefindings are not surprising if one views the pallidal output asplaying a modulatory or gating role (Bullock and Grossberg,1988), biasing the motor system toward greater or lesser activation based on the prevailing context. This idea is consistent with the hypothesis that bradykinesia in Parkinson's disease reflects scaling deficiencies (Hallett and Khoshbin, 1980;Berardelli et al., 1986; Hallett, 1993). Normally, the recognitionof a context for action would permit continuation of thataction and, possibly, its amplification. In Parkinson's disease, afailure to recognize an appropriate context would lead to therelative suppression of ongoing action.

Also of interest are two alternative kinds of failures of striatal context recognition: failure to detect a context thatshould suppress action or falsely signaling a context for action

when one has not occurred. Such failures should lead to behavior occurring outside of an appropriate context, and it hasbeen noted that basal ganglia dysfunction could lead to awide variety of such behavior (Hallett, 1993), including thetics of Tourrette's syndrome (Petersen et aI., 1993; Singer etaI., 1993), involuntary movements such as ballismus and chorea (Page et aI., 1993), derangement of volitional movement(athetosis), and the repetitive, uncontrollable thoughts or actions that characterize obsessive-compulsive disorder (Wiseand Rapoport, 1988).

In considering basal ganglia function and dysfunction, it isalso helpful to distinguish between the gating of immediatemovements, as implied in most of the above examples, andthe planning of future actions. A planning function for thecortical-basal ganglionic module is inherent in our suggestionthat cortical-thalamic loops retain information about whichcontexts have been recognized by striatal spiny neurons.From a different perspective, there has been a recent emphasis on cognitive functions of the basal ganglia, focusing on aconcept related to planning termed attentional set. Severalstudies of cognitive deficits in Parkinson's disease havestressed these patients' difficulty in shifting set (e.g., Flowersand Robertson, 1985; Brown and Marsden, 1988; Owen et aI.,1993; see also Passingham, 1993). While it is known that thereis substantial dopamine input to neocortex in primates, it remains plausible to assume that deficits in Parkinson's diseasereflect striatal dysfunction. On that assumption, the Parkinsonian's difficulty in shifting set, or planning, is also consistentwith the idea that context detection is a fundamental striatalfunction. Failure to detect a context for implementing a planwould be analogous with failure to detect a context for action. Perhaps the descending pathways of the basal gangliaare primarily involved in gating functions, whereas the ascending pathways to the frontal cortex, especially those tothe prefrontal areas, are more important in planning on thebasis of recognized context.

Cortical-Cerebellar Modules

Recognition ofAction Goals by Purkinje CellsThe emergent functions in our model of cortical-cerebellarmodules (Fig. 2A) derive from an ability of Purkinje (and basket) cells to recognize patterns that lead to the selection andexecution of actions (Berthier et aI., 1993) and from the propensity of positive feedback between the motor cortex andcerebellum to serve as a driving force for generating movement commands (Houk et aI., 1993). As discussed in the previous section, Purkinje neurons have cellular properties thatmake them ideal pattern classifiers. We posit that climbingfibers train Purkinje cells to recognize complex input patternsthat signify when the goal of an action is about to beachieved. Purkinje cell firing then inhibits cortical-cerebellarloop activity, terminating the movement command to achievean accurate movement.

Convergence onto Purkinje and Basket CellsAs with the cortical-basal ganglionic modules, the convergence of diverse inputs is important in the operation of cortica1-cerebellar modules. In this case, parallel fiber (PF) afferents,shown in Figure 2A, converge with ratios up to 200,000: 1onto Purkinje cells (PCs) and with lesser ratios onto basketcells (BCs). Parallel fibers arise from granule cells, which receive mossy-fiber inputs from a variety of sources (notshown). One source, the corticopontocerebellar projections,is similar to the corticostriatal projections in that they originate from much of the cerebral cortex (Brodal, 1987; Steinand Glickstein, 1992). Cortical columns originating these fibers are labeled "C" in Figure 2A, except for the motor col-

umn (M). As with the corticostriatal system, there is substantial convergence from these corticopontine inputs. Also likethe corticostriatal system, a certain degree of segregation ofinputs can also be found (Bjaalie and Brodal, 1989), and various combinations of frank overlap and interdigitation probably exist.

Cortical-Cerebellar Modules as Selectors andExecutors ofActionThe core of the cortical-cerebellar module is a recurrent loopthrough the frontal cortex (Houk, 1989: Houk et aI., 1993).This loop courses from one of the divisions of motor cortex(M in Fig. 2A), via the pons (not shown), to cerebellar nucleus(N) as a mossy fiber collateral (Shinoda et aI., 1992), and returns to motor cortex though the thalamic (T) link of a cortical-thalamic module. We envision that the cortical-cerebellarnetwork is actually composed of many positive-feedbackloops (Houk et aI., 1993), and when regenerative activity isdistributed throughout the network, this instantiates the motor act. Regenerative activity can be triggered by excitatoryinputs to any component or combination of components ofthe feedback loops. The modules effect their control of limbmovement through the corticospinal and rubrospinal pathways.

While superficially similar, the feedback loops involvingthe cortical-cerebellar and cortical-basal ganglionic modulesdiffer in important ways. Cortical-cerebellar loops involve predominantly excitatory transmission at each stage in the loop,which should function well in transmitting graded signals tocommand graded actions. The disinhibitory loops inherent inthe cortical-basal ganglionic modules may, instead, function ina bistable manner and be better suited for transmitting discrete pattern-recognition signals. In cortical-basal ganglionicmodules, the pattern-recognizing elements are embedded inthe main recurrent loop. In contrast, the parts of the corticalcerebellar module that operate as pattern recognizers, Purkinje and basket cells, are in a separate loop of the module. SincePurkinje cells inhibit nuclear cells (Fig. 2A), Purkinje cell discharge serves to regulate the main loop's activity.

Following the work of Marr (1969) and Albus (1971), Houkand Barto (1992) postulated that the special neuronal architecture of the cerebellar cortex is used to recognize complexstates that are critical for the selection and control of actions,as well as for translation of this state information into spatiotemporal patterns of motor discharge appropriate for commanding precise movements. Information about the state ofthe organism and its environment is conveyed to Purkinje andbasket cells by parallel fibers. The parallel fiber system, inturn, receives its information from ascending sensory, reticulocerebellar, and corticopontocerebellar pathways throughtheir mossy fiber inputs. The Purkinje and basket cells learnhow to respond to these inputs as a result of training signalsin climbing fibers that emanate from the inferior olivary complex.

The view that climbing fibers act as detectors of discoordination stems from an analysis of the physiological properties of olivary neurons (Houk and Barto, 1992; also see Kawato and Gomi, 1992). In our model, climbing fibers traincortical-cerebellar modules to move the limb from a startingposition to a selected target or endpoint in space, the lattercomprising the goal that Purkinje cells have learned to recognize. Since the modules are intrinsically capable of generating motor patterns that command straight trajectories froman arbitrary starting position to a selected goal (Berthier etaI., 1993), there is no fundamental requirement for the higherstage of trajectory formation that is assumed in most alternative models of limb control (e.g., Kawato et aI., 1987; Miallet aI., 1993). However, our model does require the assistance

Cerebral Cortex Mar/Apr 1995. V 5 N 2 99

Figure 2. A, A cortical-cerebellar module, with an embedded cortical·thalamicmodule, in the format of Figure lA. C,cortical column; M, motor cortical column; T, thalamocortical neuron; N, deepcerebellar nuclear neuron; BC, basketcell; PC, Purkinje cell; PFs, parallel fibers.B, Activation patterns ofsome ofthe slements in a cortical-cerebellar module,inthe format ofFigure 1B. Traces BC, PC,and N·T·M represent signals in basketcells, Purkinje cells, and cortical-cerebellar loops, respectively. For traces PCand N-T-M, the solid lines illustrate modules that are disinhibited by basket cellfiring, thus supporting an action command. Incontrast, thestippled lines illustrate modules with suppressed command output due to strong Purkinje cellactivity.

A

C

B

PFs

BCnselection signal

actioncommand

programmingand terminating

action

PC

N-T-M

if'instruction

Iif'

trigger

actioncommand

: :..;\,

\

J\

of reprogramming signals from a higher neural stage (presumably the premotor cortex) whenever an indirect trajectory toa target is desired, as, for example, when the movement hasto avoid an obstacle. These properties accord with the findingthat premotor lesions that spare primary motor cortex do not

100 Distributed Modules in Motor Control> Houk and Wise

interfere with ordinary reaching, while rendering an animalincapable of moving its limb around an obstacle to grasp anobject (Moll and Kuypers, 1977).

Figure 2B shows how signals for controlling a direct trajectory might be generated in a cortical-cerebellar module.

Although scant evidence exists for the activity patterns ofsome elements, such as the basket cells, we are postulatingthat a population of basket cells is tuned to respond (traceBC in Fig. 2B) to a pattern of input states that determine,along with parallel fiber inputs, which Purkinje cells will beinhibited (solid trace PC) or excited (stippled trace) prior tothe start of a movement. The change in Purkinje cell statecaused by the selection signal from the basket cells and parallel fibers constitutes the programming of an action in ourscheme. The bistability of Purkinje dendrites could functionas a temporary memory of this selection, spanning the timeinterval shown in Figure 2B between an instruction cue thatselects a motor program and a trigger cue that initiates loopactivity comprising the action command (trace N-T-M). Notethat if each of the major dendrites of a Purkinje cell werebistable, the cell as a whole could express multiple stablefiring rates, thus incorporating both increases and decreasesin discharge as illustrated by trace PC in Figure 2B.

According to our model, Purkinje cells active at the beginning of the action command will inhibit the cortical-cerebellar loops at selected sites in the cerebellar nuclei, thus determining the precise spatiotemporal motor pattern establishedin the cortical-cerebellar network. Since several hundred Purkinje cells converge upon each nuclear cell, if relatively fewof the Purkinje cells discharge at the beginning of movement,a relatively large signal can emanate from the associated loop,which presumably leads to a larger and/or faster movement.Consistent with this prediction, the discharge of task-relatedcerebellar nuclear cells correlates with the velocity and duration of the movement (van Kan et aI., 1993a). PCs showboth increases and decreases in discharge in association withmovement (Mano and Yamamoto, 1980; Chapman et aI., 1986;Dugas and Smith, 1992; Stein and Glickstein, 1992; Thach etaI., 1992; Fortier et aI., 1993). The decreases could permit thebuildup of positive feedback in modules that innervate agonist muscles, whereas the increases could suppress the activity of modules that innervate antagonists (cf. Houk et al.,1993).

It is important to note that, in this scheme, the programmed action is not initiated by the programming process;instead, it is initiated by separate triggering signals. Sensoryand cortical-cortical inputs to the cortical-cerebellar loop atany point can trigger positive feedback (Fig. 2A), thus initiating the action command. If the triggering process were suppressed, one could have programming activity in the cerebellar cortex without any execution of movement, in agreementwith the finding that local blood flow in the cerebellum isenhanced when a subject imagines a movement sequence,without actually performing it (Decety et aI., 1990).

Assuming that the triggering process does occur, as themovement proceeds the Purkinje cells that turned off duringselection would recognize that the desired action is nearlycompleted, whereupon they would resume their stable "on"state and inhibit positive feedback in the cortical-cerebellarmodule. This is considered to be a predictive type of controlthat can compensate for the time delays between a centralcommand and the resultant action. Miall et al. (1993) proposed that the cerebellum accomplishes predictive compensation by forming detailed models of the controlled systemand combining these models in a specialized manner to generate the output command. By contrast, we suggested ascheme in which Purkinje cells simply combine signals representing limb and target position, velocity, force, and efference copy signals to determine when to switch to their "on"state, thus terminating the motor command (van Kan et al.,1993b). Purkinje cells would learn to recognize input patternsthat occur at a significant phase advance so as to predict goal

acquisition rather than depending on delayed feedback confirming that a goal has actually been achieved.

Cell Activity in Different Components of the ModuleThe proposed model provides an explanation for the substantial differences that have been observed between movementrelated signals that enter and leave the cerebellum (Houk andGibson, 1987; Houk et aI., 1990; van Kan et aI., 1993b). Themost striking differences between inputs recorded frommossy fibers and outputs recorded from either Purkinje ornuclear cells concern sensory responsiveness and sensitivityto position versus velocity of a movement. Many mossy fibersare highly responsive to sensory stimulation and show prominent tonic sensitivity to limb position; they may also showsensitivity to velocity (van Kan et aI., 1993b). In contrast, theoutputs recorded from Purkinje (Harvey et aI., 1977) and nuclear (Harvey et aI., 1979; van Kan et aI., 1993a) cells showeither no response or a weak, phasic response to natural sensory stimulation, whereas they show intense movement-related bursts that correlate with velocity. These data documentstriking sensory-to-motor and position-to-velocity transformations that occur within the cerebellar cortex.

The observed transformations are well accounted for bythe proposed model of cortical-cerebellar modules. Accordingto the model, sensory signals related to limb position wouldinform Purkinje cells about the progress of ongoing movements and would contribute importantly to the constellationof inputs these cells might use to recognize when the limbis near the goal of an action. The Purkinje cells would thenswitch to an intense firing mode, which would terminate themovement command in advance of the desired endpoint ofthe movement. The bistable dendritic properties discussedearlier lead to the prediction that Purkinje (and, therefore,nuclear) cells would be relatively insensitive to sensory inputsfrom their parallel fibers under most circumstances. Onlywhen the cell neared the point of state transition would inputs appear to be effective. This aspect of the model explainsthe sensory-to-motor transformation of the cerebellar cortex.As noted earlier, the present model generates the phasic, velocity-related aspect of the output signals by shutting downthe cortical-cerebellar loop activity that was originally triggered to initiate the movement command (Fig. 2B). This feature of the model accounts for the position-to-velocity transformation of the cerebellar cortex.

In contrast to the differences predicted between cerebellarcortical inputs and outputs, the model predicts that the activity of cells in the deep cerebellar nuclei should resemble signals recorded in the primary motor cortex and at other stagesthroughout the limb premotor network, at least to a first approximation (Houk et aI., 1993). Some early reports suggestedthat activity in the deep cerebellar nuclei leads that in theprimary motor cortex (see Thach et aI., 1992). However, Fortier et al. (1993, p 1143) have recently undertaken a detailedcomparison of cerebellar and motor cortical activity duringreaching movements. They reported no dramatic timing difference between cerebellar and motor cortical premovementactivity and that, notwithstanding some important contrastsin detail, "the mean directional tuning curves of the cerebellarand motor cortical populations suggest strong similaritiesamong the general behavior of these two motor controlregions." The overall similarity of activity patterns, timing, andother properties confirms the predictions of the present model. The differences between cerebellar and neocortical activity might also be instructive. Motor cortical neurons tendedto show more inhibition of activity during nonpreferred directions of limb movement, were more narrowly tuned fordirection, and showed less variability than cerebellar neurons.There were more nondirectional cerebellar cells and they


showed a higher spontaneous firing rate than primary motorcortical neurons. In sum, it appears that the cerebellar component of the loop is more likely to be actively dischargingthan its cortical limb, at least for forearm-movement tasks, butdoes so in a manner less consistently correlated with motorbehavior. This property, which might result from a lack ofinhibitory interneurons in the cerebellar nucleus, could biasthe cortical-cerebellar loop toward its threshold for sustainingactivity.

Function and Dysfunction ofCortical-Cerebellar ModulesIn the present model, Purkinje cells act as adaptive patternclassifiers, and failures in their function will have predictableconsequences for motor control. Excessively restrictive orpermissive classifications will lead to errors in movement amplitude and direction. Coordination can be thought to resultfrom the proper control of Purkinje cells throughout the relevant parts of the cerebellar cortex and the timing of thatcontrol during the evolution of ongoing movements andmovement sequences. This model is consistent with the effects of cerebellar disease, which are characterized by errorsof starting, stopping, and directing action, as well as controlof acceleration, velocity, force, and the synthesis of complexsynergies (Holmes, 1939; Stein and Glickstein, 1992; Thach etaI., 1992).

The cerebellum has been traditionally viewed as a structure of coordination and, more recently, of motor learning(Ito, 1984: Thompson, 1986). Certainly, many structures participate in coordination and motor learning, as postulated formodels of the vestibulo-ocular reflex (Ito, 1984; Lisberger,1988; Peterson et aI., 1991) and saccadic eye movements(Dominey and Arbib, 1992; Houk et aI., 1992). But, as Thachet al. (1992, p 436) have concluded, "what is unique aboutcerebellar learning is the flexibility, ease, and speed of assigning the trigger-response ... and of building the tailor-madecomplex movement response." The architecture described inthis section, positing Purkinje cell control of distributed cortical-cerebellar loop discharge, shows one simple mechanismfor achieving those attributes. Input to any part of the loopfrom any modality, source, or level of the neuraxis can triggerthe same motor program. Various combinations of theseloops, regulated to a particular spatiotemporal patternthrough the mediation of inputs to the cerebellar cortex, cansculpt the choice of components into a coherent, coordinatedact of unending complexity and flexibility.

Cortical-Cortical and Cortical-Thalamic Modules

Collective Computations with Pyramidal CeUsIn an earlier section we suggested that cortical pyramidal cellsmay be specialized for graded input/output relations, whichwould make them well suited to the performance of collective computations (Tank and Hopfield, 1987). Collective computations involve weighting and combining several input factors to generate an output that reflects the aggregateproperties of these inputs. As opposed to the pattern classification problems faced by Purkmje and striatal neurons, themyriad of inputs to pyramidal cells would have to be addedand subtracted to form a graded resultant that reflects manyfactors over a wide, dynamic range, rather than calling forsharp decision surfaces.

Not only are neurons with graded input/output propertiesof value in collective computations; their organization intohighly interconnected, recurrent networks represents an important element in their computational capabilities. A longhistory of connectionist literature (Kohonen, 1972; Andersonet aI., 1977; Grossberg, 1980; Hopfield, 1982, 1984; Rummel-


hart et aI., 1987) has shown that highly interconnected, recurrent networks are an excellent architecture for associativememories, perceptual operations, and the solution of optimization problems. The ability to respond to graded inputs allows many individual factors to be appropriately weighted;the presence of many neurons takes advantage of the speedof parallel computation; and the availability of many loops ofinterconnection facilitates recursive readjustments of multiplefactors. These features in combination yield networks ideallysuited to automatic collective computations.

Nodal Points and Attractors. Inputs converge upon a cortical column from several different cortical areas (cortical-cortical modules), from other columns of the same cortical fields,and from several thalamic nuclei (cortical-thalamic modules)(Jones et aI., 1978; Jones, 1985; Goldman-Rakic, 1988a; Huntley and Jones, 1991). Cortical columns connecting with eachother form cortical-cortical modules and the same columnsconnecting with restricted zones in the thalamus form cortical-thalamic modules. Although cortical cells projecting toother cortical columns and those reciprocating inputs to thalamus reside in different layers, for the most part they arehighly interconnected and receive many common inputs. Thecortical column can be thought of, then, as a nodal point thatcombines the several types of influence represented by itsdiverse sources of cortical and thalamic input (Goldman-Rakic, 1988a,b).

As indicated in Figures 1 and 2, nearly all of the connections that converge upon a particular nodal point in the cerebral cortex are reciprocated by return projections, althoughlaminar organization and quantitative aspects of these recriprocal connections may differ (Goldman-Rakic, 1988a; Felleman and Van Essen, 1991). Reciprocity combines with thenodal architecture to form a complex cortical and thalamicnetwork that comprises numerous feedback loops and alternative paths for information flow. Networks of this type haveinteresting nonlinear dynamical properties that may includelimit cycles, chaotic trajectories, and discrete equilibriumstates called point attractors (Freeman, 1979; Hopfield, 1982).The focus here will be on point attractors, which are pointsin state space' where state variables remain constant ratherthan varying as a function of time. Attractors can be thoughtof as valleys, or minima, in a landscape created by an "energy"function (Hopfield, 1982). If the state of the network is transiently displaced away from an attractor, energy is elevated,and the state variables (the activations or discharge rates ofneruons) will begin to change with time, either increasing ordecreasing. Their variations will be along a trajectory thatgradually moves the state to a lower energy level, that is, toward an attractor. The network can settle back to either thesame attractor, or to another one, as long as it follows a downhill trajectory on the energy landscape. Connectionist research has demonstrated that the particular values assumedby the state variables at attractor points can be used to represent solutions to important information processing problems in cognitive psychology (Rummelhart et aI., 1987; Tankand Hopfield, 1987: Hinton and Shallice, 1991).

The feedback loops discussed earlier for cortical-cerebellarmodules can serve to illustrate the general idea of networkattractors. These recurrent loops are characterized by twotypes of point attractor (Eisenman et aI., 1991). A quiescenttype would serve to keep the network in a relatively inactivestate, corresponding to a resting period before a motor command is generated. In contrast, an active type of attractorwould maintain the network in a state of intense activity, corresponding to the population discharge that comprises a motor command. While the animal rests, network state mightvary somewhat but remain within the basin of attraction ofthe quiescent attractor, A sufficiently strong sensory input

sensory input motor output

Figure 3. The cerebral cortex is represented as an idealized, two-dimensional sheet. Nodal points may have high Iwhite and light stipple versus lowblack and heavy stipplel activityduring a sensorimotor event. Neither laminar organization nor temporal variation is reflected inthis figure.

would push the network state out of the quiescent basin andinto the basin of the active attractor. This would amount totriggering the action, as outlined above. The position of theactive attractor in state space would be controlled by thepattern of Purkinje cell discharge impinging on the corticalcerebellar loops. Particularly important would be the silencing of Purkinje cells at selective nodes in the network to permit the buildup of the intense discharge needed to commanda movement in a particular direction. A transition back intothe quiescent basin would result from the resumption of firing in these Purkinje cells, promoted by the parallel fiber inputs that occur as the goal of the movement is being approached.

To return to cortical-cortical modules, Figure 3 shematizesactive cortical nodal points (lighter areas) as viewed from thesurface of a flattened cortical map, and one can imagine thereciprocating thalamic neurons underlying each of thesenodes. This diagram should be viewed not as a snapshot at aparticular point in time, but as a time exposure with the"shutter" open throughout the duration of an entire sensorimotor response. It is also important to emphasize that theactivity map is quite different from the energy landscape thatdirects the movements of trajectories in state space, althoughthe two are related. The activity map conforms well to athree-dimensional representation, with two dimensions devoted to the idealized, flattened cortical surface, whereas network state space will have a much higher dimensionality (seenote 1). The sensory input on the left in Figure 3 is assumedto activate an initial site on the cortex. This activity thenspreads to other areas, eventually activating the motor cortexso as to trigger recurrent activity in cortical-cerebellar loops,as discussed in the previous paragraph. A connectionist inter-

pretation of this spread of activity from sensory to motorareas might relate it to a trajectory Winding downhill throughthe energy landscape characterizing n-dimensional statespace, ultimately moving the activity of the network to anattractor that signifies a particular spatial pattern of motorcortical activation.

Cerebral Cortex as a Repository of State Information. Inaddition to serving as an attractor network for collective computations, the cerebral cortex also serves as a repository formuch useful information about the state of the environmentand the organism, that is, external and internal conditions. Theactivities of neurons in columns throughout a given area ofthe cortex typically have response properties that are similarin some generic sense, whereas one or several parametric features of the responses differ among the individual columns.In this manner, an array of cortical columns, as a whole, formsa distributed representation of some internal state of the organism or an external state of the world. Examples of suchrepresentations include those of movement direction in themotor cortex (Georgopoulos et al., 1993), visual image motionin the visual cortex (Gilbert and Wiesel, 1981), and space inthe parietal and prefrontal cortex (Zipser and Anderson,1988; Funahashi et al., 1989).

Coordination of Multiple Modules in Controlling BehaviorIn the section on cortical-basal ganglionic modules, we advanced the hypothesis that they might function as detectorsof specific contexts, providing information that could be useful in the planning and gating of action. Then, in the sectionon cortical-cerebellar modules, we discussed how they mightfunction in the programming, execution and termination ofactions. Finally, in the section immediately above, we indicated


contextregistration

BASAL GANGLIA

sensory inputs

CEREBELLUM

actioncommand

Figure 4. Combined schema ofacortical-basal ganglionic and acortical-cerebellar module to illustrate how activities inmultiple modules could be coordinated (see text). Abbreviationsare as in Figures lA and 2A.

how cortical-cortical modules might implement collectivecomputations while also serving as a repository for diversedistributed representations. While it is clear that each of theseoperations might contribute usefully to the control of action,an important problem concerns how different modules canfunction cooperatively. In this section, we posit two levels ofcoordination among modules: one within arrays of similarmodules, the other among different kinds of modular arrays.

Coordination within ArraysAt the most elementary level of modular organization, we envision arrays of similar modules that must function in a coordinated manner to control the population activity in eachcortical area. We do not illustrate a modular array, but theycan be readily imagined as n replications of Figure ZA for thecortical-cerebellar class of modules and n repetitions of Figure lA for the cortical-basal ganglionic modules. The clearneed for coordination within these arrays can be appreciatedfrom the well-known coding of single-unit activity in the motor cortex (Georgopoulos et al., 1993). Cortical neurons arebroadly tuned to movement direction, and a population ofneurons with different optimal tuning directions needs to berecruited and coordinated to command a particular limbmovement. Simulations have shown (Berthier et al., 1993)that an array of cortical-cerebellar modules can learn to function in a coordinated manner under the training influence ofsimulated climbing fiber inputs to the cerebellar cortex. Theinformation source for coordination in this particular modelis feedback from the environment and other types of inputtransmitted via the mossy fiber-parallel fiber input to Purkinjecells, as envisioned by Thach et al. (1992).


Coordination between individual modules may also resultfrom divergent anatomical connections among functionallyrelated loops (Houk et aI., 1993), which could promote therecruitment of functionally related modules near the beginning of movement and the simultaneous cessation of discharge at its end. Thus, coordination within arrays of corticalcerebellar modules may stem from either the adaptive use ofcommon inputs, interconnections between functionally related modules, or both. Coordination within arrays of corticalbasal ganglionic, cortical-thalamic, and cortical-cortical modules is probably achieved through the application of similarprinciples.

Coordination among ArraysAt a higher level of organization, the motor system needsmechanisms for coordinating the functions of different modular arrays. Three general types of anatomical projections maypromote communication and coordination among arrays: (I)corticostriatal, (2) corticopontocerebellar, and (3) corticocortical projections. Figure 4 illustrates examples of these typesof communication between one cortical-basal ganglionicmodule, one cortical-cerebellar module, and a group of cortical-cortical and cortical-thalamic modules. If one imagineseach of these modules being replaced by an array of n similarmodules, the same diagram can be used to contemplate communication among modular arrays.

Figure 4 was constructed by joining Figures lA and ZA,using as tether points common connections with cortical columns. We assumed that the rightmost cortical column in thecortical-basal ganglionic module of Figure lA is the same asthe leftmost column in the cortical-cerebellar module of Fig-

oryes

Cerebral CortexBasal Ganglia (collective computation)

(context encoders) (information arrays)

;;7 ~~~((]bB\~ motor

m((]![j]ILl~B\~((]W'~ commandsdopamine fibers %!!(reinforcement)

Cerebellum sensclimbing fibers (pattern generators) stat

(discoordination)

Figure 5. Overview diagram showing global interrelations in a system based on a distributed modular architecture. The cerebral cortex performs collective computations thatautomate planning and acting, while also serving as a repository of information arrays used by the basal ganglia and cerebellum. The basal ganglia use this information to encodenew contexts that are fed back into frontal cortical arrays, a process that is guided byreinforcement inputs transmitted in dopamine fibers. The cerebellum uses similar informationto select and generate motor patterns, a process guided bydiscoordination signals transmitted in climbing fibers. These motor patterns are fed back to motor cortical arrays, andcorticospinal fibers transmit them to the spinal cord as motor commands. Global modulators may help to coordinate the activities of these different parts of the brain throughgeneralized reinforcement.

ure ZA, and is the middle column depicted in Figure 4. Thisillustrates the fact that many cortical columns originate bothcorticostriatal and corticopontocerebellar projections, although they originate from different neurons within the column (Mercier et aI., 1990). The differential laminar organization of corticostriatal versus corticopontine cells suggeststhat, to an extent, a cortical column could send differing signals to each, but the overall columnar organization of thecortex predicts that these signals should have much in common. The common cortical column mentioned earlier is depicted as receiving sensory information (via the thalamus),which may be useful both in the detection of a critical context and in the execution of a particular action. Hence, weillustrate that column as feeding information into both thebasal ganglia and the cerebellum. Similarly, we assumed thatthe leftmost cortical column in the cortical-basal ganglionicmodule projects additionally to the cerebellum by way of thepons. As the output of that cortical-basal ganglionic module,this leftmost column registers the fact that the context towhich the module is tuned just occurred. This informationcan be used by the cortical-cerebellar modules, thus assistingin the programming and execution of an appropriate actionbased on a recognized context. The third type of anatomicalconnectivity available for coordination is corticocortical, and,since these projections are generally reciprocal, they form thecortical-cortical information processing modules discussedabove. The rapid, automatic communication provided by cortical-cortical modules provides an efficient mechanism forsharing related information, reorganizing it, and promoting coordination among arrays of distributed modules.

Operating PrinciplesNote that each of the processes controlled by individual modular arrays may occur asynchronously and in parallel. For example, context information may be available for selecting anappropriate action before it is initiated, or alternatively, a default action can be initiated and the selection process canfunction later to redirect the commanded action as it evolves(Houk et aI., 1993). This would accommodate the independent actions observed for "where" and "when" systems ineye-limb coordination (Gielen and van Gisbergen, 1990) aswell as the asynchronous interactions observed between limbmovement initiation and its programming (Ghez et aI., 1990).

Figure 5 summarizes the global interrelations in a systembased on the distributed modular architecture discussed in

this article. Because the cerebral cortex is interposed betweenthe basal ganglia and cerebellum, one of its functions will beto serve as a switchboard for information flow between thesetwo subcortical structures. Considering the large number offunctionally distinct cortical areas, each comprising an arrayof cortical columns, one can think of the cerebral cortex asan expansive repository for a diverse set of information arrays.Some of these arrays are trained by local learning rules torepresent coarse coding of relatively unmodified sensory information, whereas another group may form more complexfeature maps based on a reorganization of primary sensoryinformation (Barlow, 1989). Yet others are controlled in morecomplex ways by input from the cerebellum and basal ganglia, via their thalamic relays. The cerebral cortex continuously uses its diverse information arrays to perform its collectivecomputations, while simultaneously sharing these data withthe basal ganglia and cerebellum.

As illustrated in Figure 5, the basal ganglia function as detectors and encoders of salient contexts. Striatal spiny neurons are trained by their dopamine reinforcement input torecognize contexts and states that are likely to be useful inguiding behavior. The inputs for these computations comefrom nearly the entire cerebral cortex, and the processed outputs return to the frontal cortex. These signals invest corticalarrays with planning information that can then be used byother modules. The cerebellum functions as a generator ofspatiotemporal patterns. To do this, Purkinje and basket cells,under the training influence of climbing fibers that detectdiscoordination, learn to select appropriate actions and to recognize the endpoints (or goals) of the actions. In addition tocommanding actions, information about intention may besent from cerebellum to motor cortical arrays, whereupon itbecomes available for use by other information processingmodules.

We envision that much of the learning that goes on in thisdistributed modular network may be module specific, due tothe topographic specificity of climbing and dopamine fiberinput to the cerebellum and basal ganglia, respectively. Whilethe specificity of dopaminergic inputs to striatum (Graybiel,1991; Gerfen, 1992) may not be quite as pronounced as thatof climbing fibers to cerebellum, both inputs are clearly topographically organized. Learning in the cortical network appears instead to depend mainly on a local Hebbian learningrule, although global modulations and reinforcements are


probably provided by diffuse neuromodulatory input to thecortex (Fig. 5).

Of particular interest from the standpoint of coordinationis the possibility that learning in the frontal cortex may beguided by the outputs from the basal ganglia and cerebellum.These outputs can be thought of as forcing functions forboth execution and learning. These forcing functions couldtrain the cortical-cortical network through the mechanism of"example setting" that was discussed in the section above oncell properties. This is because basal ganglia and cerebellarinputs to their respective cortical-thalamic modules are wellpositioned to alter the attractors of the cortical-cortical andcortical-thalamic arrays of the frontal lobe. On one hand, thesealterations would serve to modify the collective computationsbeing performed by the cortical network on a moment-bymoment basis. On the other hand, the same alterations wouldpromote changes in the weights of the Hebbian synapses onpyramidal neurons that, in the long run, would move the network's attractors closer to the points being forced by cerebellar and basal ganglia modifications. As a consequence, thefrontal cortex would become trained to perform, in a highlyefficient and automatic fashion, those particular functions being forced on it by its subcortical inputs. This hypothesis contrasts with the traditional assumption that the cerebral cortexlearns first and then trains subcortical structures.

The Search for Intelligent BehaviorHow might the cortical-basal ganglionic, cortical-cerebellar,and cortical-cortical modular arrays discussed in this articlecontribute to intelligent motor behavior? Three decades agoMinsky (1963), in his thoughtful analysis of the major unresolved problems in the field of artificial intelligence, enumerated five broad areas of problem-solving activity: search, planning, pattern recognition, learning, and induction. Searchcomprises the immediate problem of selecting a solution(e.g., an action) from a large repertoire of potential solutionsdistributed in neural space. Exhaustive search of the entiresolution space is a straightforward, but usually impractical,solution. As an alternative to such an inefficient approach,Minsky saw planning as an administrative facility for dividingcomplex problems into simpler subproblems and pattern recognition as a means to improve the efficiency of searchingfor a solution by categorizing the search into small domains.Minsky viewed learning as a means for directing, throughreinforcement, a search toward solutions that have workedfor similar problems in the past. Finally, induction can be construed as the invention of novel solutions, plans, and patternrecognition capabilities to solve problems that the individualhas never encountered. In this section, we use the global summary of Figure 5 as a basis for discussing how the basal ganglia, cerebellum, and cerebral cortex might function togetherin the implementation of Minsky's concepts for intelligent behavior.

We posit that, ultimately, the search for an action dependson cortical-cerebellar modules, taking particular advantage ofthe unique cellular specializations of Purkinje cells for patternclassification. Some automatic actions, of course, bypass thecerebellum and are controlled exclusively by brainstem andspinal premotor networks, as a reflex or a sensorimotor habit.However, when the choice of an action or its implementationis reasonably complex, the cerebellum is likely to be involved(Ito, 1984; Houk and Barto, 1992; Thach et aI., 1992). According to the theory espoused here, cortical-cerebellar arrays notonly retrieve a stored motor program, but also see throughthe execution of the action to the point of recognizing theconditions for its termination. In doing so, cortical-cerebellarmodules create an output pattern that is temporal as well asspatial and comprises the activity in a large array of effector


premotor circuits. When activated and executed, these processes and their sequelae represent the completed search.

Planning may depend on the context recognition function of the cortical-basal ganglionic modules. Through the mediation of the frontal cortex (Fig. 5), context informationcould be used by cortical-cerebellar modules to plan actionson the basis of the motivational, sensorimotor, and cognitivecontext. In an overly simplistic example, one can envision thebasal ganglia as recognizing one of two potential contexts,each associated with half of the potential solution space,which thereby speeds search by a factor of 2. By recognizinga greater variety of contexts, the basal ganglia's contributioncould promote the efficiency of the search many fold. Indeed,without such context recognition, the ability of the cerebellarcortex to use its pattern classification mechanism for choosing and implementing appropriate actions would be quitelimited (Rosenblatt, 1962; Minsky and Papert, 1969). By tapping into the cortical information arrays, the cerebellar cortexis able to draw on inputs beyond crude sensory and motorstates so as to include the wealth of processed informationavailable in the cortical arrays.

The mediation of the frontal cortex in this coordinationdeserves further comment. As discussed earlier, one type ofrecoding involves the formation of feature maps based oninformation-maximizing principles (Barlow, 1989). Although apattern classifier functions better when presented with features, an even more substantial improvement can be achievedif the pattern classifier is presented with high-order properties, provided these properties are of "heuristic" value (Minsky, 1963). The striatum is ideally situated to perform this heuristic classification, because its pattern recognition is guidedby training signals representing predictions of reward(Schultz et aI., 1995). Its spiny neurons sample a diverse inputspace that comprises sensory states, features, motor intentions, and high-level properties generated by its own contextencoding actions. The postulated ability of the basal gangliato use contexts that it initially detects to encode yet morecomplex contexts is potentially very powerful, due to the recursive nature of this computation.

To Minsky (1963), learning controls the navigationthrough solution space and does so on the basis of experience. He discussed in some detail the problem of credit assignment, which is one of the major obstacles to effectivelearning in complex systems. Credit assignment involves getting the right training information to the right location (spatial credit assignment) at the right time (temporal credit assignment) for it to be effective in guiding the learningprocess. The topographical nature of climbing fiber input tothe cerebellum (Oscarsson, 1980; Voogd and Bigare, 1980)and of dopamine projections onto the striatum (Graybiel,1991; Gerfen, 1992) could help to resolve the spatial creditassignment problem, and the predictive nature of dopaminesignals would help to resolve the temporal credit assignmentproblem (Houk et aI., 1995).

Induction represents the highest of Minsky's processes:the generation and evaluation of novel solutions representsthe central problem of intelligent behavior, We have alreadydiscussed the importance of recursion in the encoding ofever-higher level contexts. Recursion and self-reference, werethese processes to take into account the overall success orfailure of the action-guidance system as a whole, might serveas a powerful tool in the process of induction.

We conclude with a thought about thinking. By focusingon the motor system we have been able to avoid the issue ofconscious awareness, since intelligent guidance of action regularly occurs in the absence of consciousness (Goodale et aI.,1991). However, the principles of distributed modular architectures that we have disussed in this article may have rele-

vance for a broad scope of cognition, as evidenced by recenttreatises that have discussed modularity in the highest brainfunctions (Fodor, 1983; Minsky, 1986).

Notes1. The relevant "state space" is n dimensional, where n is the number of neurons and each axis characterizes a neuron's state, for example, its activation or firing rate.

This work was supported by a Center Grant from the NationalInstitutes of Mental Health (P50 MH 48185) to Dr. Houk.

Correspondence should be addressed to Dr. J. C. Houk, Department of Physiology, Northwestern University Medical Center, 303 EastChicago Avenue, Chicago, IL 60611.

ReferencesAlbin RL, Young AB, Penney JB (1989) The functional anatomy of

basal ganglia disorders. Trends Neurosci 12:366-375.Albus JS (1971) A theory of cerebellar functions. Math Biosci 10:25

61.Alexander GE (1987) Selective neuronal discharge in monkey puta

men reflects intended direction of planned limb movements. ExpBrain Res 67:623-634.

Alexander GE, Crutcher M (1990) Preparation for movement: neuralrepresentations of intended direction in three motor areas of themonkey.] NeurophysioI64:133-150.

Alexander GE, DeLong MR, Strick PL (1986) Parallel organization offunctionally segregated circuits linking basal ganglia and cortex.Annu Rev Neurosci 9:357-381.

Alexander GE, Crutcher MD, DeLong MR (1990) Basal ganglia-thalamocortical circuits: parallel substrates for motor oculomotor "prefrontal" and "limbic" functions. Prog. Brain Res 85:119-146.

Alexander ME, Wickens JR (1993) Analysis of striatal dynamics: theexistence of two modes of behavior.J Theor Bioi 163:413-438.

Allen GI, Tsukahara N (1974) Cerebro-cerebellar communication systems. Physiol Rev 54:957-1006.

Anderson JA, Silverstein]W; Ritz SA,Jones RS (1977) Distinctive features, categorical perception, and probability learning: some applications of a neural model. Psychol Rev 84:413-451.

Anderson ME, Turner RS (1991) Activities of neurons in cerebellarreceiving and pallidal-receiving areas of the thalamus of the behaving monkey.] NeurophysioI66:879-893.

Apicella p, Scarnati E, Schultz W (1991) Tonically discharging neurons of monkey striatum respond to preparatory and rewardingstimuli. Exp Brain Res 84:672-675.

Arbib, MA (1987) Brains, machines, and mathematics, 2d ed. NewYork: Springer.

Aronin N, Chase K, DiFiglia M (1986) Glutamic acid decarboxylaseand enkephalin immunoreactive axon terminals in the rat neostriatum synapse with striatonigral neurons. Brain Res 421:151158.

Asanuma C, Thach WT, Jones EG (1983) Distribution of cerebellarterminations and their relation to other afferent terminations inthe ventral lateral region of the monkey. Brain Res Rev 5:237265.

Barlow HB (1989) Unsupervised learning. Neural Comput 1:295311.

Bassant MH, Ennouri K, Lamour Y (1990) Effects of iontophoreticallyapplied monoamines on somatosensory cortical neurons of unanesthetized rats. Neuroscience 39:431-439.

Bear MF, Kleinschmitt A, Gu Q, Singer WJ (1990) Disruption of experience-dependent synaptic modifications in striate cortex byinfusion of an NMDA receptor antagonist.J Neurosci 10:909-925.

Berardelli A, Dick JPR, Rothwell JC, Day BL,Marsden CD (1986) Scaling of the size of the first agonist EMG burst during rapid wristmovements in patients with Parkinson's disease. J Neurol Neurosurg Psychiatr 49:1273-1279.

Berthier NE, Singh Sp' Barto AG, Houk JC (1993) Distributed representation of limb motor programs in arrays of adjustable patterngenerators.J Cognit Neurosci 5:56-78.

Bjaalie JG, Brodal P (1989) Visual pathways to the cerebellum: segregation in the pontine nuclei of terminal fields from differentvisual cortical areas in the cat. Neuroscience 29:95-107.

Brodal P (1987) Organization of cerebropontocerebellar connections as studied with anterograde and retrograde transport of

HRP-WGA in the cat. In: New concepts in cerebellar neurobiology(KingJS, ed), pp 151-182. New York: Liss.

Brotchie p,lansek R, Horne MK (1991) Motor function of the globuspallidus. 2. Cognitive aspects of movement and phasic neuronalactivity. Brain 114:1685-1702.

Brown RG, Marsden CD (1988) Internal vs. external cues and thecontrol of attention in Parkinson's disease. Brain 111:323-345.

Bullock D, Grossberg S (1988) Neural dynamics of planned armmovements: emergent invariants and speed-accuracy propertiesduring trajectory formation. Psychol Rev 95:49-90.

Chapman CE, Spidalieri G, Lamarre Y (1986) Activity of dentate neurons during arm movements triggered by visual, auditory, and somesthetic stimuli in the monkey.] Neurophysiol 55:203-226.

Chevalier G, Deniau JM (1990) Disinhibition as a basic process inthe expression of striatal functions. Trends Neurosci 13:277-280.

Decety J, Sjoeholm H, Ryding E, Stenberg G, Ingvar DH (1990) Thecerebellum participates in mental activity: tomographic measurements of regional cerebral blood flow. Brain Res 535:313-317.

DeFelipe J, Conley M, Jones EG (1986) Long-range focal collateralization ofaxons arising from corticocortical cells in the monkeysensory-motor cortex.J Neurosci 6:3749-3766.

DeLong MR (1990) Primate models of movement disorders of basalganglia origin. Trends Neurosci 13:281-285.

DeLong MR, Georgopoulos AP (1981) Motor functions of the basalganglia. In: Handbook of physiology, Sec 1, The nervous system,Vol II, Motor control, 2 (Brookhart JM, Mountcastle VB, Brooks VB,Geiger SR, eds), pp 1017-1061. Bethesda, MD: American Physiological Society.

Dominey PF,Arbib MA (1992) A cortico-subcortical model for generation of spatially accurate sequential saccades. Cereb Cortex 2:153-175.

Dugas C, Smith AM (1992) Responses of cerebellar Purkinje cells to

slip of a hand-held object.J NeurophysioI67:483-495.Eisenman LN, Keifer J, Houk JC (1991) Positive feedback in the cere

bra-cerebellar recurrent network may explain rotation of population vectors. In: Analysis and modelling of neural systems (Eeckman F,ed), pp 371-376. London: Kluwer.

Ekerot C-F (1984) Climbing fibre actions of Purkinje cells-s-plateaupotentials and long-lasting depression of parallel fibre responses.In: Cerebellar functions (Bloedel J, et al., eds), pp 268-274. Berlin:Springer.

Felleman DJ, Van Essen DC (1991) Distributed hierarchical processing in the primate cerebral cortex. Cereb Cortex 1:1-47.

Flaharty AW, Graybiel AM (1991) Corticostriatal transformations inthe primate somatosensory system. Projections from physiologically mapped body-part reprcscntations.] Neurophysiol 66: 12491263.

Flaharty AW, Graybiel AM (1993) Output architecture of the primateputamen.J Neurosci 13:3222-3237.

Flowers KA, Robertson C (1985) The effect of Parkinson's diseaseon the ability to maintain mental set.J Neurol Neurosurg Psychiatr48:517-529.

Fodor JA (1983) The modularity of mind. Cambridge, MA:MIT Press.Fortier PA, Smith AM, Kalaska JF (1993) Comparison of cerebellar

and motor cortex activity during reaching: directional tuning andresponse variability.J NeurophysioI69:1136-1149.

Freeman W (1979) Nonlinear dynamics of paleocortex manifestedin the olfactory EEG. Bioi Cybern 3:21-37.

Frith CD (1992) The cognitive neuropsychology of schizophrenia.Hove, UK: Erlbaum.

Funahashi S, Bruce C], Goldman-Rakic PS (1989) Mnemonic codingof visual space in the monkey's dorsolateral prefrontal cortex. JNeurophysioI61:331-349.

Fuster JM, Alexander GE (1973) Firing changes in cells of the nucleus medialis dorsalis associated with delayed response behavior.Brain Res 61:79-91.

Gardiner Tw, Nelson RJ (1992) Striatal neuronal activity during theinitiation and execution of hand movements made in response tovisual and vibratory cues. Exp Brain Res 92:15-26.

Georgopoulos Ap, Taira M, Lukashin A (1993) Cognitive neurophysiology of the motor cortex. Science 260:47-52.

Gerfen CR (1992) The neocortex mosaic: multiple levels of compartmental organization in the basal ganglia. Annu Rev Neurosci15:285-320.

Gerfen CR, Engber TM, Mahan LC, Susel Z, Chase TN, Monsma FJ,


Silbey DR (1990) 0, and 0, dopamine receptor-regulated geneexpression of striatonigral and striatopallidal neurons. Science250: 1429-1432.

Ghez C, Hening \v, Favilla M (1990) Parallel interacting channels inthe initiation and specification of motor response features. In: Attention and performance, Vol XIII, Motor representation and control (Jeannerod M, ed), pp 265-293. Hillsdale, N]: Erlbaum.

Ghosh S, Murray GM, Turman AB, Rowe M] (1994) Corticothalamicinfluences on transmission of tactile information in the ventroposterolateral thalamus of the cat: effect of reversible inactivationof somatosensory cortical areas I and II. Exp Brain Res 100:276286.

Gibson AR, Robinson FR, Alam], Houk]C (1987) Somatotopic alignment between climbing fiber input and nuclear output of theintermediate cerebellum.] Comp Neurol 260:362-377.

Gielen CCAM, van Gisbergen]AM (1990) The visual guidance ofsaccades and fast aiming movements. News Physiol Sci 5:58-63.

Gilbert C, Wiesel TN (1981) Laminar specialization and intracorticalconnections in cat primary visual cortex. In: The organization ofthe cerebral cortex (Schmitt FO, Worden FG, Adelman G, DennisSG, eds), pp 163-191. Cambridge, MA: MIT Press.

Gilbert CD, Wiesel TN (1983) Clustered intrinsic connections in catvisual cortex.] Neurosci 3:1116-1133.

Goldman PS, Nauta W]H (1977) Columnar distribution of corticocortical fibers in the frontal association, limbic, and motor cortexof the developing rhesus monkey. Brain Res 122:393-413.

Goldman-Rakic PS (1984) Modular organization of prefrontal cortex.Trends Neurosci 7:419-429.

Goldman-Rakic PS (1988a) Topography of cognition: parallel distributed networks in primate association cortex. Annu Rev Neurosci11:137- I 56.

Goldman-Rakic PS (1988b) Changing concepts of cortical connectivity: parallel distributed cortical networks. In: Neurobiology ofneocortex (Rakic P, Singer \v, eds), pp 177-202. New York: Wiley.

Goldman-Rakic PS, Friedman HR (1991) The circuitry of workingmemory revealed by anatomy and metabolic imaging. In: Frontallobe function and dysfunction, Chap 4 (Levin HS, Eisenberg HM,Benton AL, eds), pp 72-91. New York: Oxford UP.

Goodale MA, Milner AD,]akobson LS, Carey DP (1991) A neurological dissociation between perceiving objects and grasping them.Nature 349: 154-156.

Graybiel AM (1991) Basal ganglia-input, neural activity, and relationto the cortex. Curr Opin Neurobiol 1:644-651.

Grossberg S (1980) How does a brain build a cognitive code? PsychRev 87:1-51.

Grossberg S, Kuperstein M (1989) Neural dynamics of adaptive sensory-motor control. New York: Pergamon.

Groves PM (1983) A theory of the functional organization of theneostriatum and the neostriatal control of voluntary movement.Brain Res Rev 5:109-132.

Hallett M (1993) Physiology of basal ganglia disorders: an overview-Silversides lecture. Can] Neurol Sci 20:177-183.

Hallett M, Khoshbin S (1980) A physiological mechanism of bradykinesia. Brain 103:301-314.

Harvey R], Porter R, Rawson]A (1977) The natural discharges ofPurkinje cells in paravermal regions of lobules V and VI of themonkey's cerebellum.] Physiol (Lond) 271:515-536.

Harvey R], Porter R, Rawson]A (1979) Discharges of intracerebellarnuclear cells in monkeys.] Physiol (Lond) 297:559-580.

Hazrati L-N,Parent A (1992) Convergence of subthalamic and striatalefferents at pallidal level in primates: an anterograde double-labeling study with biocytin and PHA-L.Brain Res 569:336-340.

Hikosaka ° (1989) Role of basal ganglia in initiation of voluntarymovements. In: Dynamic interactions in neural networks: modelsand data (Arbib MA,Amari S, eds),pp 153-167. New York: Springer.

Hikosaka 0, Wurtz RH (1983) Visual and oculomotor functions ofmonkey substantia nigra pars reticulata. IV Relation of substantianigra to superior colliculus.] NeurophysioI49:1285-1301.

Hikosaka 0, Sakamoto M, Usui S (1989a) Functional properties ofmonkey caudate neurons. II. Visual and auditory responses.] NeurophysioI61:799-813.

Hikosaka 0, Sakamoto M, Usui S (1989b) Functional properties ofmonkey caudate neurons. III. Activities related to expectation oftarget and reward.] NeurophysioI61:814-832.

108 Distributed Modules in Motor Control- Houk and Wise

Hinton GE, Shallice T (1991) Lesioning an attractor network: investigations of acquired dyslexia. Psychol Rev 98:74-95.

Holmes G (1939) The HugWings Jackson memorial lecture, The cerebellum of man. Brain 62:1-30.

Hoover ]E, Strick PL (1993) Multiple output channels in the basalganglia. Science 259:819-821.

Hopfield]] (1982) Neural networks and physical systems with emergent collective computational abilities. Proc Natl Acad Sci USA79:2554-2558.

Hopfield]] (1984) Neurons with graded response have collectivecomputational properties like those of two-state neurons. ProcNatl Acad Sci USA 81:3088-3092.

Horak FB, Anderson ME (1984) Influence of globus pallidus on armmovements in monkeys. I. Effects of kainic acid-induced lesions.]Neurophysiol 52:290-304.

Houk]C (1989) Cooperative control of limb movements by the motor cortex brainstem and cerebellum. In: Models of brain function(Cotterill RM], ed), pp 309-325. Cambridge: Cambridge UP.

Houk ]C, Barto AG (1992) Distributed sensorimotor learning. In: Tutorials in motor behavior II (Stelmach GE, Requin], eds), pp 71100. Amsterdam: Elsevier.

Houk ]C, Gibson AR (1987) Sensorimotor processing through thecerebellum. In: New concepts in cerebellar neurobiology (King]S, ed), pp 387-416. New York: Liss.

Houk ]C, Wise SP (1993) Outline for a theory of motor behavior. In:Research notes in neural computing, Vol IV; Neuroscience: fromneural networks to artificial intelligence (Rudomin P, Arbib MA,Cervantes-Perez F, eds), pp 452-470. Heidelberg: Springer.

Houk ]C, Singh Sp' Fisher C, Barto AG (1990) An adaptive sensorimotor network inspired by the anatomy and physiology of thecerebellum. In: Neural networks for control (Miller WT, Sutton RS,Werbos P], eds), pp 301-348. Cambridge, MA:MIT Press.

Houk ]C, Galiana HL, Guitton 0 (1992) Cooperative control of gazeby the superior colliculus, brainstem and cerebellum. In: Tutorialsin motor behavior II (Stelmach GE, Requin], eds), pp 443-474.Amsterdam: Elsevier.

Houk ]C, Keifer ], Barto AG (1993) Distributed motor commands inthe limb premotor network. Trends Neurosci 16:27-33.

Houk ]C, Adams ]L, Barto AG (1995) A model of how the basalganglia generates and uses neural signals that predict reinforcement. In: Models of information processing in the basal ganglia,Chap 13 (Houk ]C, Davis ]L, Beiser DG, eds), in press. Cambridge,MA: MIT Press.

Huntley G\v,]ones EG (1991) Relationship of intrinsic connectionsto forelimb movement representations in monkey motor cortex:a correlative anatomic and physiological study.] Neurophysiol66:390-413.

Ito M (1984) The cerebellum and neural control. New York: Raven.Jones EG (1985) The thalamus. New York: Plenum.Jones EG, Coulter ]0, Hendry SHC (1978) Intracortical connections

of architectonic fields in the somatic sensory, motor, and parietalcortex of monkeys.] Comp Neurol 181:291-348.

Kawaguchi Y, Wilson C], Emson PC (1990) Projection subtypes ofrat neostriatal cells revealed by intracellular injection of biocytin.] Neurosci 10:3421-3438.

Kawato M, Gomi H (1992) A computational model of four regionsof the cerebellum based on feedback-error learning. BioI Cybern68:95-103.

Kawato M, Furukawa K, Suzuki R (1987) A hierarchical neural-network model for control and learning of voluntary movement. BioICybern 56:1-17.

Kimura M (1990) Behaviorally contingent property of movementrelated activity of the primate putamen.] NeurophysioI63:12771296.

Kimura M (1992) Behavioral modulation of sensory responses ofprimate putamen neurons. Brain Res 578:204-214.

Kimura M, Rajkowski], Evarts EV (1984) Tonically discharging putamen neurons exhibit set-dependent responses. Proc Natl AcadSci USA 81:4998-5001.

Kimura M, Kato M, Shimazaki H (1990) Physiological properties ofprojection neurons in the monkey striatum to the globus pallidus.Exp Brain Res 82:672-676.

Kohonen T (1972) Correlation matrix memories. IEEETrans ComputC-21:353-359.

Kornhuber HH (1971) Motor functions of cerebellum and basal gan-

glia: the cerebellocortical saccadic (ballistic) clock, the cerebellonuclear hold regulator, and the basal ganglia ramp (voluntaryspeed smooth movement) generator. Kybernetik 8:157-162.

Lisberger SG (1988) The neural basis for learning of simple motorskills. Science 242:728-735.

Ljungberg T, Apicella l' Schultz W (1991) Responses of monkey dopamine neurons to external stimuli: changes with learning. In: Thebasal ganglia 111 (Bernardi G, Carpenter MB, Di Chiara G, MorelliM, Stanzione l' eds), pp 487-494. New York: Plenum.

Ljungberg T, Apicella P Schultz W (1992) Responses of monkey dopamine neurons during learning of behavioral reactions.] NeurophysioI67:145-163.

Ulnas R (1988) The intrinsic electrophysiological properties ofmammalian neurons: insights into central nervous system function. Science 242:1654-1664.

Llinas R, Sugimori M (1980) Electrophysiological properties of invitro Purkinje cell dendrites in mammalian cerebellar slices. ]Physiol (Lond) 305:197-213.

Mano N, Yamamoto K (1980) Simple-spike activity of cerebellar Purkinje cells related to visually guided wrist tracking movement inthe monkey.] Neurophysiol 43:713-728.

Marr D (1969) A theory of cerebellar cortex.] Physiol (Lond) 202:437-470.

Marsden CD (1982) The mysterious motor function of the basalganglia: the Robert Wartenburge lecture. Neurology 32:514-539.

McCormick DA, Connors BW; Lighthall ;W; Prince DA (1985) Comparative electrophysiology of pyramidal and sparsely spiny stellateneurons of the neocortex.] Neurophysiol 54:782-806.

Mercier B, Legg C, Glickstein M (1990) Basal ganglia and cerebellumreceive different somatosensory information in rats. Proc NatlAcad Sci USA 87:4388-4392.

Miall RC, Weir D], Wolpert DM, Stein]F (1993) Is the cerebellum aSmith predictor? ] Mot Behav 25:203-216.

Mink]W; Thach WT (1991) Basal ganglia motor control. III. Pallidalablation: normal reaction time muscle cocontraction and slowmovement.] NeurophysioI65:330-351.

Minsky ML (1963) Steps toward artificial intelligence. In: Computersand thought (Feigenbaum EA, Feldman ], eds), pp 406-450. NewYork: McGraw-Hill.

Minsky M (1986) The society of mind. New York: Simon and Schuster.

Minsky ML, Papert SA (1969) Perceptrons. Cambridge, MA: MITPress.

Moll L, Kuypers HG]M (1977) Premotor cortical ablations in monkeys: contralateral changes in visually guided reaching behavior.Science 198:317-319.

Mountcastle VB (1978) An organizing principle for cerebral function: the unit module and the distributed system. In: The mindfulbrain (Edelman GM, Mountcastle VB, eds), pp 36-41. Cambridge,MA: MIT Press.

Murphy PC, Grieve KL, Sillito AM (1993) Effects of vasoactive intestinal polypeptide on the response properties of cells in area 17of the cat visual cortex.] NeurophysioI69:1465-1474.

Nilsson N] (1965) Learning machines. New York: McGraw-Hill.Oscarsson 0 (1980) Functional organization of olivary projections

to the cerebellar anterior lobe. In: The inferior olivary nucleus:anatomy and physiology (Courville], Montigny CD, Lamarre, eds),pp 279-289. New York: Raven.

Owen AM, Roberts AC, Hodges ]R, Summers BA, Polkey CE, RobbinsTW (1993) Contrasting mechanisms of impaired attentional setshifing in patients with frontal lobe damage or Parkinson's disease.Brain 116:1159-1175.

Page RD, Sambrook MA, Crossman AR (1993) Thalamotomy for thealleviation of levodopa-induced dyskinesia: experimental studiesin the I-methyl-4-phenyl-l ,2,3,6-tetrahydropyridine-treated Parkinsonian monkey. Neuroscience 55:147-165.

Parthasarathy HB, Schall ]D, Graybiel AM (1992) Distributed but convergent ordering of corticostriatal projections: analysis of the frontal eye field and the supplementary eye field in the macaque monkey.] Neurosci 12:4468-4488.

Passingham RE (1993) The frontal lobes and voluntary action. NewYork: Oxford UP.

Pennartz CMA, Kitai ST (1991) Hippocampal inputs to identifiedneurons in an in vitro slice preparation of the rat nucleus accum-

bens-evidence for feed-forward inhibition.] Neurosci 11:28382847.

Peterson BW;Baker ]F, Houk]C (1991) A model of adaptive controlof vestibuloocular reflex based on properties of cross-axis adaptation. Ann NY Acad Sci 627:319-337.

Petersen B, Riddle MA,Cohen D], Katz LD, Smith ]C, Hardin MT,Leekman]F (1993) Reduced basal ganglia volumes in Tourette's syndrome using three-dimensional reconstruction techniques frommagnetic resonance images. Neurology 43:941-949.

Rakic P (1988) Specification of cerebral cortical areas. Science 241:170-176.

Reiter HO, Striker MP (1989) Neural plasticity without postsynapticaction potentials: less active inputs become dominant when visualcortical cells are pharmacologically inhibited. Proc Natl Acad SciUSA 85:3623-3627.

Rolls ET,Thorpe S],Madison SP (1983) Responses of striatal neuronsin the behaving monkey. 1. Head of the caudate nucleus. BehavBrain Res 7:179-210.

Romo R, Schultz W (1992) Role of primate basal ganglia and frontalcortex in the internal generation of movements. III. Neuronal activity in the supplementary motor area. Exp Brain Res 91:396407.

Rosenblatt F (1962) Principles of neurodynamics. New York: Spartan.

Rummelhart DE, Hinton GE, McClelland]L (1987) A general framework for parallel distributed processing. In: Parallel distributedprocessing, Vol, 1, Foundations (Rummelhart DE, McClelland ]L,eds), pp 45-76. Cambridge, MA: MIT Press.

Schlagger BR, Fox K, O'Leary DDM (1993) Postsynaptic control ofplasticity in developing somatosensory cortex. Nature 364:623626.

Schultz W; Romo R (1988) Neuronal activity in the monkey striatumduring the initiation of movements. Exp Brain Res 71:431-436.

Schultz W; Romo R (1992) Role of primate basal ganglia and frontalcortex in the internal generation of movements. I. Preparatoryactivity in the anterior striatum. Exp Brain Res 91:363-384.

Schultz W;Apicella P Ljungberg T (1993) Responses of monkey dopamine neurons to reward and conditioned stimuli during successive steps of learning a delayed response task.] Neurosci 13:900-913.

Schultz W; Romo R, Ljungberg T, Mirenowicz ], Hollerman]R (1995)Dopamine neurons carry reward-related signals. In: Models of information processing in the basal ganglia, Chap 12 (Houk ]C, Davis ]L, Beiser DG, eds), in press. Cambridge, MA: MIT Press.

Selemon L, Goldman-Rakic P (1985) Longitudinal topography andinterdigitation of corticostriatal projects in the rhesus monkey.]Neurosci 5:776-794.

Selemon L, Goldman-Rakic P (1988) Common cortical and subcortical targets of the dorsolateral prefrontal and posterior parietalcortices in the rhesus monkey: evidence for a distributed neuralnetwork subserving spatially guided behavior.] Neurosci 8:40494068.

Shallice T (1998) From neuropsychology to mental structure. Cambridge: Cambridge UP.

Shepherd GM (1990) The synaptic organization of the brain, 3d ed.London: Oxford UP.

Shinoda Y,Kakei S (1989) Distribution of terminals of thalamocortical fibers originating from the ventrolateral nucleus of the catthalamus. Neurosci Lett 96:163-167.

Shinoda Y, Sugiuchi Y, Futami T, Izawa R (1992) Axon collaterals ofmossy fibers from the pontine nucleus in the cerebellar dentatenucleus.] NeurophysioI67:547-560.

Shinoda Y, Kakei S, Futami T, Wannier T (1993) Thalamocorticalorganization in cerebello-thalamo-cortical system. Cereb Cortex 3:421-429.

Singer HS, Reiss AL, Brown ]E, Aylward EH, Shih B, Chee E, Harris EL,Reader M], Chage GA, Bryan RN, DenckJa MB (1993) VolumetricMRI changes in basal ganglia of children with Tourette's syndrome. Neurology 43:950-956.

Stein]F, Glickstein M (1992) The role of the cerebellum in the visualguidance of movement. Physiol Rev 72:967-1017.

Tank DW; Hopfield]] (1987) Collective computation in neuron-likecircuits. Sci Am 255(12):104-114.

Thach WT, Goodkin HG, Keating]G (1992) Cerebellum and the


adaptive coordination of movement. Annu Rev Neurosci 15:403442.

Thompson RF (1986) The neurobiology of learning and memory.Science 233:941-947.

Tremblay L, Filion M (1989) Responses of pallidal neurons to striatalstimulation in intact waking monkeys. Brain Res 498: 1-16.

van Kan PLE, Houk JC, Gibson AR (1993a) Output organization ofintermediate cerebellum of the monkey. J Neurophysiol 69:5773.

van Kan PLE, Gibson AR, Houk JC (1993b) Movement-related inputsto intermediate cerebellum of the monkey.] NeurophysioI69:7494.

Voogd J, Bigare F (1980) Topographical distribution of olivary andcorticonuclear fibers in the cerebellum. A review. In: The inferiorolivary nucleus (Courville J, de Montigny C, Lamarre Y, eds), pp207-234. New York: Raven.

Wang X-J, Rinzel J, Rogawski MA (1991) A model of the T-type calcium current and the low-threshold spike in thalamic neurons. JNeurophysiol 66:839-850.

Waterhouse BD, Azizi SA, Burne RA, Woodward OJ (1990) Modulation of rat cortical area 17 neuronal responses to moving visualstimuli during norepinephrine and serotonin microiontophoresis.Brain Res 514:276-292.

Wickens J (1990) Striatal dopamine in motor activation and rewardmediated learning: steps towards a unifying model. J NeuralTransm 80:9-31.

Wilson CJ (1990) Basal ganglia. In: The synaptic organization of thebrain (Shepherd GM, ed), pp 279-316. London: Oxford UP.

Wilson CJ (1995) The contribution of cortical neurons to the firingpattern of striatal spiny neurons. In: Models of information processing in the basal ganglia, Chap 3 (Houk JC, Davis JL, Beiser DG,eds), in press. Cambridge, MA: MIT Press.

Wilson CJ, Groves PM (1980) Fine structure and synaptic connections of the common spiny neurons of the rat neostriatum: a studyemploying intracellular injections of horseradish peroxidase. JComp Neurol 194:599-615.

Wise SP, Houk JC (1994) Modular neuronal architecture for planningand controlling behavior. Bioi Commun Danish R Acad Sci Lett43:21-33.

Wise Sp,RapoportJL (1988) Obsessive-compulsive disorder: is it basal ganglia dysfunction? In: Obsessive-compulsive disorder in children and adolescents (Rapoport IL, ed), pp 327-344. Washington,DC: American Pychiatric.

Yeterian EH, Van Hoesen GW (1978) Cortico-striate projections inthe rhesus monkey: the organization of certain cortico-caudateconnections. Brain Res 139:43-63.

Yuen GL, Hockberger PE, Massone LE, Houk JC (1992) Qualitativedynamical model of bistability in Purkinje cell dendrites. Soc Neurosci Abstr 18:1206.

Zipser 0, Anderson RE (1988) A back-propagation programmed network that simulates response properties of a subset of posteriorparietal neurons. Nature 331 :679-684.


Distributed Modular Architectures Linking Basal...

Documents

Transcript of Distributed Modular Architectures Linking Basal...