Motor Control & Sensory Integration

Obituary Remembering Denis Glencross I first met Denis in June of 1979 at a NATO Advanced Study Institute on Motor Learning and Control held at a former Benedictine abbey in Senanque in France. Them, over glasses of an incredibly purple vin ordinaire, and in walks among the heather, we debated the concept of the motor schema and I came to appreciate both Denis's enthusiastic, warm personality, and his broad interests in skilled performance, whether in sport, in carefully designed experiments, or in conceptual analysis. The breadth of these interests is attested by the fact that at that time (in the years 1979 and 1980) he published in the Journal of Motor Behavior, Neuropsychologica, Sportswissenschaften, and Brain and Language as well as in volumes on psychology of Motor Behavior and Sport, Tutorials in Motor Behavior (the record of the Senanque meeting), and Attention and Performance VIII. Denis John Glencross was born in Perth on February 15, 1939, and gained his Bachelor's and Master's degrees at the University of Western Australia in 1961 and 1964, where he also served as Lecturer in Physical Education. He then moved somewhat East, to Adelaide where he remained - with scholarly leaves in London, Cambridge, Eugene in Oregon, Marseilles, and Wassenaar in the Netherlands - until 1988. He studied at the University of Adelaide where he received his PhD in Psychology in 1972, At the same time, he rose through the ranks at Flinders University, as Lecturer, Senior Lecturer and then Associate Professor. It was during his time at Flinders that I twice met Denis again. Living in the United States, with parents in Sydney and in-laws in Perth, I had the perfect excuse for a day's stop-over in Adelaide which combined intense discussion on issues in motor control with the search for the perfect glass of port in MacLaren Vale. These visits not only reinforced my high opinion of Denis's own work and of the broad intellectual enthusiasm he brought to it, but also gave me a sense of how well he worked with his colleagues and students.

vi Obituary The Flinders years saw the development of the field of sports psychology - his influential book Sport and Psychology was published in 1978 - together with a succession of papers on a broad range of topics, including careful laboratory experiments with normal subjects, analysis of keyboard skills, and insightful analysis of a variety of special populations, including Parkinson's disease patients, clumsy children and injured athletes. In 1988, Denis at last had the chance to return to Perth, as Professor of Psychology and Head of the School of Psychology at Curtin University of Technology. There, Denis soon built up an enthusiastic group with whom he continued to explore a wide range of detailed experiments, and scholarly issues, as well as addressing such applied topics as training young drivers and the effects of drug use on human performance. Through all this time, Denis played an active role in Australian professional societies, including the arrangement of diverse symposia and workshops. It was at one of these that I last met Denis - at the \"Motor Control and Human Skills Research Workshop\" held in the seaside town of Mandurah, south of Perth in December of 1993 - the meeting on which the present volume is based. The meeting brought together staff (faculty, as they say in the States) and students from all around Australia. The talks and posters were overall of great interest, and I was particularly impressed by the work of young physiotherapists which ably combined practice and theory. Denis was in fine, enthusiastic form, and provided energetic leadership to the meeting, including a catalytic role in chairing a discussion session which helped tie the many themes of the meeting together, and acting as master of ceremonies for the conference banquet, presenting awards he had chosen himself to provide humorous recognition of the contributions different people had made to the success of the meeting. As I worked on my paper for this volume, then, it was with this image of Denis's intellectual and personal energy and enthusiasm very much in mind. Yet, eight months later, on August 18, 1994, he was dead. It is still hard to believe, and leaves a void in the lives of those who knew him, and in the study of motor control and sports psychology in Australia and overseas. Michael A. Arbib University of Southern California February 1995

vii Preface This volume evolved from the Second Biennial Motor Control & Human Skill Research Workshop which was held in Mandurah, Western Australia, in December of 1993. This meeting involved Australian and international speakers interested in motor control and the broader problems of serial organisation and sensory-motor integration of human skills. A number of specific issues were highlighted, including the planning and programming of action, amendment and up-dating mechanisms, and in particular, perception-action coupling, coordination and adaptation, neural mechanisms and the representation of action. Underlying much of the session was the major theoretical issues which included the debate between computational and prescriptive approaches versus the emergent properties and system dynamics approaches. This volume represents a diverse approach from disciplines such as psychology, electrical and mechanical engineering, human movement studies, physiotherapy, neurology, and kinesiology. Such a broad range of interests provides many different perspectives to the understanding of sensory-motor integration. I wish to express my appreciation to all the contributors for their efforts, not only for producing the high quality chapters presented in this volume, but also for their assistance in the reviewing process. In addition, I wish to thank Nick Barrett, Peter Livesey, Andrea Lamont-Mills, David Livesey, Moyra Tsouvallas, Phillip Bairstow and Gary Thickbroom for their helpful comments on several chapters in the volume. I would like to acknowledge Rosemary Skinner for her assistance with the indexes, and Jeanette Vaughan, who had the unenviable task of collating the chapters into one congruous volume. She has endless patience. I am particularly grateful to Michael Arbib, Bruce Abemethy, Jeff Summers and John Warm, who always found time to provide support and encouragement following Denis's death. I wish to dedicate this book to the memory of Denis Glencross. The inception of this volume can be attributed to the initiative and foresight of Denis, who, in 1991, decided to run a Workshop dedicated to research on motor control and human skill. This provided a unique opportunity for both researchers and students working in the area to exchange ideas and converse at a more personal level. This proved so successful that it was continued as a biennial event. Thank you Denis for your friendship, your support and your brilliance. Jan P. Piek March, 1995

xi Contributors BRUCE ABERNETHY Department of Human Movement Studies University of Queensland ST LUCIA Qld 4072 Australia MICHAEL A. ARBIB Center for Neural Engineering University of Southern California Los Angeles CALIFORNIA 90089-2520 USA NICHOLAS C. BARRETI' School of Psychology Curtin University of Technology GPO Box U 1987 PERTH 6000 Western Australia PATRICIA BATE Department of Physiotherapy La Trobe University BUNDOORA Victoria 3053 Australia JOHN BRADSHAW Department of Psychology Monash University CLAYTON Victoria 3168 Australia ROBIN BURGESS-LIMERICK Department of Human Movement Studies University of Queensland ST LUCIA Queensland 4072 Australia RICHARD G. CARSON School of Kinesiology Simon Fraser University BURNABY, BC, V5A 156

xii List of Contributors ROSS CUNNINGTON Department of Psychology Monash University CLAYTON Victoria 3168 Australia DIGBY ELLIOTr Department of Kinesiology McMaster University HAMILTON ON, L8S 4K1 Canada CRAIG ENGSTROM Department of Human Movement Studies University of Queensland ST LUCIA, Qld 4072 Australia DENIS J. GLENCROSS School of Psychology Curtin University of Technology GPO Box U 1987 PERTH 6000 Western Australia DAVID GOODMAN School of Kinesiology Simon Fraser University BURNABY, BC, V5A 1S6 Canada ALASTAIR HANNA Department of Human Movement Studies University of Queensland ST LUCIA, QLD 4072 Australia ERROL HOFFMANN Department of Mechanical & Manufacturing Engineering Melbourne University PARKVILLE Victoria 3052 Australia ROBERT IANSEK 3192 Geriatric Research Unit, Kingston Centre Warrigan Road CHELTENHAM Victoria Australia

List of Contributors xiii ROBERT KANE Department of Psychology University of WA NEDLANDS WA 6009 Australia J.A. SCOTT KELSO Center for Complex Systems Florida Atlantic University 500 NW Street BOCA RATON FL 33431, USA GRAHAM K. KERR University Laboratory of Physiology Oxford University Parks Road OXFORD OX1 3PT England BOB N. MARSHALL Department of Human Movement & Recreational Studies University of WA NEDLANDS WA 6009 Australia THOMAS MATYAS Department of Behavioural Health Sciences La Trobe University BUNDOORA Victoria 3083 Australia MEG E. MORRIS Schools of Physiotherapy & Behavioural Health Sciences La Trobe University BUNDOORA Victoria 3083 Australia ROBERT NEAL Department of Human Movement Studies University of Queensland ST LUCIA, QLD 4072 Australia

xiv List of Contributors MEGAN NEILSON Clinical Research Unit for Anxiety Disorders University of New South Wales At St Vincent's Hospital 299 Forbes Street DARLINGHURST NSW 2010 Australia PETER NEILSON School of Electrical Engineering University of New South Wales SYDNEY NSW 2052 Australia NICHOLAS O'DWYER Cerebral Palsy Research Unit Institute of Neurological Sciences The Prince Henry Hospital LI'IqI,E BAY NSW 2036 Australia JAMES PHILLIPS Department of Psychology Monash University CLAYTON Victoria 3168 Australia JAN P. PIEK School of Psychology Curtin University of Technology GPO Box U 1987 PERTH WA 6001 Australia JEFF PRESSING 3052 Department of Psychology University of Melbourne PARKVILLE Victoria Australia SIMON K. RUSHTON Department of Psychology University of Edinburgh 7 George Square EDINBURGH Scotland

List of Contributors XV NICOLAS SCHWEIGHOFER Center for Neural Engineering University of Southern California LOS ANGELES, CA 90089-2520 USA ANDRAS SEMJEN Laboratorie de Neurosciences Cognitives CNRS MARSEILLE, France JEFFERY J. SUMMERS Department of Psychology University of Southern Queensland PO Darling Heights TOOWOOMBA Qld 4350 Australia W.T. THACH Washington University School of Medicine Department of Anatomy and Neurobiology ST LOUIS MO 63110-1031 USA JULIE THOMAS Department of Psychology University of Southern Queensland TOOWOOMBA Qld 4350 Australia JOHN P. WANN Department of Psychology University of Edinburgh SCOTLAND EH8 952

Motor Control and Sensory Motor Integration: Issues and Directions D.J. Gleneross and J.P. Piek (Editors) 9 1995 Elsevier Science B.V. All fights reserved. Chapter I MOTOR CONTROL AND SENSORY-MOTOR INTEGRATION Denis J. Glencross School of Psychology, Curtin University of Technology In this chapter, sensory-motor integration is described as a cyclical process of perception-action coupling. Several different perspectives on sensory-motor integration are presented, including hierarchical neurophysiological and dynamical systems approach. A mixed model approach to understanding motor control is discussed. Sensory-motor integration, the theme of the present book, is concerned with the interaction and integration of sensory factors in the on-going organisational control of motor processes and hence movements and actions. This definition, however, continues to extend the 'artificial' dichotomy between sensory processes and motor or output operations. For whatever reason, the distinction between sensory and motor functions has neglected the very essence of the problem, and that is, sensory and motor, or alternatively, perception-action-perception coupling, is a cyclic phenomenon. The cycle itself cannot be fully understood if any one component is 'surgically removed' and studied in isolation. Sensory-motor integration really means that the system is tightly coupled and interdependent at all levels, and that this is both the advantage of such a system, but at the same time a 'disadvantage' of the system. The notion of perception-action-perception coupling raises a second important issue, namely, it is this very feature which characterises all human skills, ranging from speech and musical skills at one extreme, to many industrial, commercial and sporting skills at the other. Much of the experimental research over the last half a century has often trivialized the complexity of human skill by studying simple tasks often in artificial or meaningless context, in which 'skilled' is defined by several hundred This manuscriptwas incomplete at the timeof DenisGlencross's death.The introductionwas includedin this volumeas it introducedthe main themeof the book,namelysensory-motor integration. The abstract was not part of the originalmanuscript.

4 D.J. Glencross practice trials, instead of the millions of trials which characterises the expert's performance (Crossman, 1959). A related concern has been the artificial segregation of much of the research endeavor into two somewhat isolated camps, one pre-occupied with demonstrating the on-going sensory contribution and control of movements and actions, and the other describing the prescriptive and planned control of action in advance of action in the minutist detail- the so-called peripheral-centralist debate (Glencross, 1977; Lashley, 1951). Not only has such a distinction been limited in its overall understanding of the problem at hand, but 'either-or' mentality has restricted the perspective on the larger and arguably more important (and interesting) question, and that is, the interaction or coupling between the sensory and motor performance. Indeed, the very designation of the fields in which many of us work, as motor control, motor learning and motor skill serve to extend 'the lie'. Certainly one can mtionalise (if not justify) the isolation of a small field of study to control and study it systematically. However, the consequence of such an approach is that an expanding and elaborating scientific community sees this narrow field as the context and the problem, and loses sight of the broader context. Indeed, the motor control and motor learning domain may be seen as an example of this paradigmatic narrowing, and I believe, to the detriment of the broader field of study of sensory-motor integration and human skill. It may be timely to revisit Karl Lashley through his seminar, the 1951 Hixon Symposium lecture, and recall his concerns about the problem of serial order which characterised all complex skills from the songs of birds, to the trotting horse, to the architect and carpenter, and of course, most noticeably in speech and language. Lashley's immediate concern was about integration in perception and action: \"Temporal integration is not found exclusively in language; the co-ordination of leg movements in insects, the song of birds, the control of trotting and pacing in a gaited horse, the rat running the maze, the architect designing a house, and the carpenter sawing a board present a problem of sequences of action which cannot be explained in terms of successions of external stimuli.\" (Lashley, 1951, p. 113). Lashley (1951) addressed the central issue of the problems of serial order or serial organisation and how this is represented in the brain. Lashley regarded the serial order problem as central to the understanding of humans' unique ability to learn new

Motor Control and Sensory-Motor Integration complex sequences of behaviour as in speaking, playing musical instruments, typing and the like. The essence of such sequences is that a relatively small number of units or elements need to be re-arranged, re-ordered, re-organised into relatively stable, coherent and meaningful sequences. Thus, the serial order problem needs to be addressed at several levels: i. the nature of the elementary units and the elementary movement worked. ii. the second level of analysis concerns the representation of the sequence of action (viz., the serial order and internal timing or phasing of such sequences). iii. the flexibility and adaptability issue needs to be more fully addressed to accommodate how semi-rigid programs can be amended to contextual variation. In contrast to the hierarchical view of Lashley (1951), Wiesendanger (1990) argued that from the neurophysiological perspective, there is evidence that many structures at all levels of the neuraxis 'may' operate in parallel, with many interconnections and interactions, and with many opportunities for back-propagation. 'Parallel distributed processing' represents more appropriately the organisation of perception and movement plans, as these \"are not encoded in single sets of neurons, but in widely distributed, multiple sets, each encoding different aspects.\" (p. 72). Further, there is strong evidence (in cortical motor areas) for much of the processing to occur in parallel as well as serial processing. However, Wiesendanger concurs that \"if hierarchies are seen mainly as levels of increasing complexity and of increasing abstraction of the sensory representation and of the motor commands, then the principle of hierarchical organisation might still be a fruitful concept\" (p. 71). Neurophysiology plays an important role in cognitive science in providing hypotheses about fruitful architectures for human cognition (e.g., human intelligence and machine intelligence). Furthermore, neurophysiology has provided the lead for the elaboration of connectionist approaches and specifically connectionist architectures. An alternative theoretical approach to the representational or computational approach has been designated the dynamical approach (Kelso, 1988; Kugler, 1986). It has been the criticisms leveled at the computational approach by the systems dynamics approaches which have brought to the fore the question of sensory-motor integration. In dynamic self-organising systems, one of the basic tenets is that perception and action are tightly coupled. This approach proposes that movement organisation is a

6 D.J. Glencross consequence of self-organising system dynamics. One consequence of this debate has been to imply that only one approach is correct - however, this seems to be an untenable and indeed undesirable theoretical stance. A solution to the degrees of freedom problem is to avoid computations by relying on inherent dynamical characteristics of the effector (limb), namely the task dynamics framework. The basic idea of task dynamics is to formulate movement problems in an abstract task-based coordinate system in which the equations governing movement are simple and uncoupled (e.g., mass spring dynamics) and characterised by low- dimension attractors (e.g., target end-point). A major virtue of task dynamics is that many of the details of trajectory planning and responding to perturbation arise directly from the structure of the dynamical equation and need not be dealt with explicitly at higher levels. \"Higher levels simply instantiate an invariant dynamical organization in the task space.\" (Jordan & Rosenbaum, 1989, p. 735). Jordan (1990) proposed a model which integrates the notions of a generalised motor program (computational) and a dynamic systems component. The theoretical issue at stake is NOT whether advance planning or advance programming occurs, but rather what details of the up-coming action can be planned in advance, and how this planning and preparation is influenced by and responds to concurrent changes in sensory information. This is the question of how updating and amendment of any advance preparation achieved within the context and time- constraints of the on-going action. Thus, sensory-motor integration driven by the degrees of freedom problem implies that there is no hard and fast distinction between organisation of the sensory and motor systems and that indeed perception and action are tightly coupled. In much the same way, it is artificial to talk about separate 'central' and 'peripheral' explanation mechanisms of motor control. The former invoke computation and centrally prescribed formulations or plans which were separate and separated from direct control by sensory information arising essentially in the periphery. Rather one might refer to what Sheridan (1991) called 'mixed models' of motor control in which these peripheral and central mechanisms are closely integrated in the organisation and control of action and human skill. One of the intriguing features of highly skilled performance, in dance for example, or in playing a symphony orchestra, is that perception and action are time-locked and

Motor Controland Sensory-MotorIntegration tightly coupled. This also occurs in all situations, even where the perceptual processing is not as obvious as in the case of the dancer or musician. The tennis player, pilot and lathe operator couple their skilled actions to a complex sensory array which to the expert is highly structured and patterned. Sensory-motor integration is the centre piece of perception-action coupling. REFERENCES Crossman, E.R.F.W. (1959). A theory of the acquisition of speed skill. Ergonomics, 2, 153-166. Glencross, D.J. (1977). Control of skilled movements. Psychological Bulletin, 84, 14- 29. Jordan, M. (1990). Motor learning and the degrees of freedom problem. In M. Jeannerod (Ed.) Attention and Performance Xlll:Motor Representation and Control (pp. 796-836). Hillsdale, NJ: Lawrence Erlbaum & Assoc Inc. Jordan, M., & Rosenbaum, D.A. (1989). Action. In Michael I. Posner (Ed.) Foundations of Cognitive Science (pp. 727-767), Cambridge, MA: MIT Press. Kelso, J.A.S. (1988). Introductory remarks: Dynamic patterns. In J.A.S. Kelso, A.J. Mandell & M.F. Shlesinger (Eds.) Dynamic patterns in complex systems (pp. 1-5). Singapore: World Scienctific. Kugler, P.N. (1986). A morphological perspective on the origin and evolution of human movement patterns. In M.G. Wade & H.T.A. Whiting (Eds.) Motor development in children: Aspects of coordination and control. Dordrecht: Martins Nijhoff. Lashley, K. S. (1951). The problem of serial order in behavior. In L.A. Jeffress (Ed.) Cerebral Mechanisms in Behavior (pp. 112-136). New York: Wiley. Sheridan, T.B. (1991). Telerobofics, automation, and human supervisory control. Cambridge: MIT Press. Wiesendanger, M. (1990). The motor cortical areas and the problem of hierarchies. In M. Jeannerod (Ed.) Attention and Performance Xlll:Motor Representation and Control (pp.59-75). Hillsdale, NJ: Lawrence Erlbaum & Assoc Inc.

Motor Control and Sensory Motor Integration: Issues and Directions D.J. Glencross and J.P. Piek (Editors) 9 1995 Elsevier Science B.V. All rights reserved. Chapter 2 MODELING THE CEREBELLUM: FROM ADAPTATION TO COORDINATION Michael A. Arbib and Nicolas Schweighofer Centerfor Neural Engineering, University of Southern California Los Angeles, CA 90089-2520 W.T. Thach Washington University School of Medicine, Department of Anatomy and Neurobiology, St. Louis MO 63110-1031 We review data showing that the cerebellum is required for adaptation both of saccadic eye movements to consistent shifts in target position and of throwing when the subject wears a wedge prism. We then model the saccade adaptation in terms of plasticity of synapses from parallel fibers to Purkinje cells in cerebellar cortex, stressing the integration of cerebellar cortex and nuclei in microzones as the units for correction of motor pattern generators. The model uses a \"window of eligibility\" to ensure that error signals that elicit a corrective movement are used to adjust the original movement, not the secondary movement. We also find that correction involves not only adjustment of the original motor pattern generator (MPG) but modulated deployment of other MPGs to yield a successful overall movement. Finally, we extend this model to account for adaptation of throwing. 1. THE ROLE OF C E R E B E L L U M IN ADAPTATION The cerebellum has been implicated in adaptation of the metrics of movement to changing circumstances. In this section, we review two examples of adaptation - - for saccades and throwing - - and briefly note evidence for the role of the cerebellum. In later sections, we will develop a model for this role, arguing that the cerebellum \"works\" by modulating and coordinating multiple Motor Pattern Generators (MPGs).

12 M.A. Arbib, N. Schweighofer & W.T. Thach 1.1 Saccade Adaptation Saccades are very fast eye movements of very short duration. As pointed out by Robinson (1986), the visual feedback delays are longer (about 40-80ms) than the movement itself (on the order of 50ms); and so saccades cannot be controlled by a normal feedback controller for accurately locating a visual target in the fovea. In a target perturbation experiment, a non-trained monkey (Goldberg et al., 1993) or a human subject (Albano and King, 1989) has to make a saccadic eye movement towards a target. During the saccade, the target is shifted to a new position but this shift is not perceived by the subject during the m o v e m e n t - we speak of \"saccadic suppression\". As the first saccade does not end at the new target position, it appears incorrect, and a second, corrective, saccade is generated with a latency comparable to the latency of the first (Albano and King, 1989). [In fact, small errors which are nunified by a single following corrective saccade appear to be a part of the normal human or monkey strategy (Optican, 1982). The reason is that the brainstem saccade generator uses a \"noisy integrator\", so completion of the saccade does not guarantee that the eye is on target.] However, over a few hundred trials, the amplitude and direction of the initial saccade changes and the amplitude of the corrective saccade decreases until the trained animal can saccade directly to the displaced target. The gain changes gradually and recovers gradually, and the gain for similar directions and amplitudes is also changed (Goldberg et al., 1993). The learning curve shows an exponential time course for the adaptation, with recovery apparently faster than learning (Figure la). The influence of the cerebellum and its associated structures on the execution of saccades can be observed in cerebellar patients and monkeys. Ritchie (1976) made symmetrical lesions of lobes VI and VII and found that large saccades made toward the primary position (centripetal) were grossly hypermetric, while those made away from primary position (centrifugal) were hypometric. Goldberg et al., (1993) studied monkeys with interpositus and fastigial nuclei lesions. Figure l b shows the lack of learning for the target perturbation experiment for monkeys after lesions of the interpositus and fastigial nuclei. The \"learning\" curve is a straight line (apart from noise). Moreover, due to the loss of the modulation supplied by the cerebellum, the saccades have a greater amplitude than those before the adaptation runs in the normal monkey and the performance, taken as the variance around the mean curve, is poorer after than before the lesion. Thus, this results suggest that the adaptation occurs in a system which includes the cerebellum and that the performance is somewhat degraded by cerebellar lesions. As noted by Ito (1984), the contribution of the vermis may be to modify central command signals executing a saccade, and Noda et al., (1990) showed that the cerebellum is indeed not the primary domain of the signal processing. Cerebellar impulses are projected downstream to saccade-programming circuits where visual

Modeling the Cerebellum: From Adaptation to Coordination 13 information has already been converted into motor-commanding signals. The cerebeUar eye movement map does not provide the total saccade command for a given frontocollicular eye movement command, but rather the correction that \"modulates\" the command issued by the superior colliculus and other regions in response to the retinal input. Prelesion Postlesion t~ 99 . : :\" . 9 9: 9 . . o 9 9 9 . .. . . ~ E :: :): O3 200 400 600 I ii l I9 Trial Number t43 0 200 400 600 0 Trial Number Figure 1. Effect of eerebellar lesion on saecadic adaptation. Each dot is a single trial. Continuous line is a ten trial running average of saccade amplitude plotted against the trial number of the middle of the epoch. (a) Normal monkey. Dashed line is at end of adaptation runs, after which the target is no longer displaced. Co) Monkey with lesion of interpositus and fastigial nuclei. There is no adaptation and performance is degraded. (From Goldberg et al. 1993). This supports the hypothesis that the cerebellum adjusts an MPG rather than being the MPG. Below, we shall further argue that coordination of MPGs is also required for successful saccade adaptation. But first we turn to a second data set, that on adaptation of throwing, to provide a broader challenge for our cerebellar modeling. 1.2 Prism Adaptation of Throwing Martin et al. (in press), in Thach's laboratory, have studied the adjustment of eye, head (gaze), ann and hand in humans throwing a dart or ball at a target while wearing wedge prism spectacles. In throwing, the eyes (and head) fixate the target, and serve as reference aim for the arm. If wedge prism spectacles are placed over the eyes with the base to the fight, the optic path is bent to the subject's fight, and the eyes (and head)

14 M.A. Arbib, N. Schweighofer & W.T. Thach move to the left in order to see the target so that the arm, calibrated to the line of sight, will throw to the left of the target. But with repeated throws, the calibration will change, and the arm will throw closer to and finally on-target. When, after adaptation, the prisms are first removed, the eyes are now on-target, but the eye-head-arm calibration for the previously left-bent gaze persists: the arm throws to the right of target by an amount almost equal to the original leftward error. With repeated throws, eye-head position and ann synergy are recalibrated: each throw moves closer to and finally on targeL Figure 2 shows the distance of the hit location to left or right of a target before donning the prisms (before the first dashed line), while wearing the prisms (between the dashed lines), and after removing the prisms (after the second dashed line). The initial points are all relatively close to the center of the target; the middle points (wearing prisms) start to the left of center, while the latter points (after prisms) start equally far to the fight. The failure to hit the target center after removal of the lenses comes as a surprise to all subjects in this study, and so it- and, we infer, the original adaptation - is unlikely to be due to a voluntary strategy. Operationally we attribute it to some subconscious adaptation. Nevertheless, conscious strategic corrections - \"cheating\" - are possible. One subject, after donning the prisms, noted that the first throw hit approximately 40 cm to the left of center but his next throw was fight on target! After being told \"Quit cheating! Throw to where you see the target, not to where you think the target actually is\", his performance then followed the adaptation curve. Another subject had seen from the performance of prior subjects the error which the prisms introduced upon the first throw. She decided she would compensate for this by aiming at a point on the opposite side of the target. But unbeknownst to her, she was given a set of prisms with base to the left rather than to the right. Upon donning them, she threw with a doubled error. Is the adaptation visual and global, or motor and specific for the trained body parts, or somewhere in between? To address this question, Martin et al. (in press) asked subjects to make both fight hand and left hand throws, with and without prisms, to see if there was carry-over of adaptation on the one task to the other task. They found that prism adaptation occurred in the throwing arm, did not affect or abate with throws by the other arm, and readapted only during throws by the first arm. Therefore the adaptation could not be considered to be generally of vision, but instead to be to some extent specific for the trained body parts. Given this result, how specific is the adaptation to the task? Does the adaptation of the trained body parts carry over to their use in other tasks, or is it specific for the use of those body parts during the one task only? To address this question, we asked subjects to use the same ann to make underhand and overhand throws, with and without prism adaptation.

Modeling the Cerebellum: From Adaptation to Coordination 15 cm Throw Figure 2. Performance of a human throwing a ball at a target with and without prisms. The vertical axis of the graph shows the distance of the hit location to left or fight of a target in a series of trials before donning the prisms (up to first dashed line), while wearing the prisms (up to second dashed line), and after removing the prisms. On wearing prisms, all subjects adapted the overhand throw. For two subjects, the subsequent underhand throw showed absolutely no effect of prior overhand adaptation.

16 M.A. Arbib, N. Schweighofer & W.T. Thach In these subjects, prior overhand adaptation persisted in subsequent overhand throws despite intervening underhand throws, and readapted with repeated overhand throws. In 4 subjects, the results were similar, except for an apparent carry-over of the overhand prism adaptation to the first subsequent underhand throw. Nevertheless, this disappeared with the second throw, and it is therefore unclear to what extent this represented an adaptive change. In these as in the first 2 subjects, the prior overhand adaptation survived the intervening underhand throws, persisting undiminished in subsequent overhand throws, and readapting only after repeated overhand throws. Two subjects showed persistent carryover from prior overhand adaptation to underhand throws, but only one showed carryover of overhand adaptation to the underhand throws which then readapted without any apparent adaptation left in the final overhand throw. To test if one can learn to store more than the one gaze-throw calibration simultaneously, Martin et al. asked subjects to make 200 throws while wearing the prisms and 250 without each day, 4 days per week for 7 weeks. They measured the progress on the 5th day of each week .with 25 throws before, 100 throws during, and 75 throws after wearing the prisms. This made a total of 900 throws with prisms and 1100 throws without prisms each week for 7 weeks. Over time and practice, the first throw with the prisms landed closer to the target, and the first throw without the prisms (aftereffect) also landed closer to the target. By 7 weeks, throws were on-target for the first trial wearing and the first trial after removing the \"known\" prisms. This suggests that 2 adaptations (no-prisms and known-prisms) may be stored' simultaneously and separately. A subject who had adapted to one prism behaved as if naive when presented with a novel prism. Both non-prism and \"known\" prism calibrations were affected; both had to be readapted independently. Prism adaptation in macaques is abolished by cerebeUar lesion (Baizer & Glickstein, 1974). Weiner et al. (1983) gave more detailed results in patients with cerebeUar disease, and showed that adaptation was not impaired by disease of corticospinal or basal ganglia systems. Martin et al. (in press) also applied their paradigm to patients with cerebellar disease. In a patient who had multiple sclerosis, with tremor and ataxia (no other deficits), no adaptation was seen after donning and doffing the prisms. With a patient with fight cerebellar hemisphere infarct, tremor and ataxia, little adaptation is seen after donning and doff'rag the prisms. Martin et al. also found that two patients with MRI- documented inferior olive hypertrophy (a degenerative disease of the inferior olive, which is the exclusive source of the cerebellar climbing fibers) could not adapt, despite otherwise normal performance. Both patients had ataxia of gait (damage of the inferior olive leads immediately to malfunction and ultimately to atrophy of the cerebellum [Strata, 1987; Murphy and O'Leary, 1971] ), but the upper extremity movements were relatively normal. This suggests that the adaptation mechanism could be dissociated at least in degree from those of coordination and performance. Finally, they studied

Modeling the Cerebellum: From Adaptation to Coordination 17 patients with lesions presumed to involve mossy fibers (see the next section for a description of cerebellar input pathways) of the middle cerebellar peduncle, who also show impaired adaptation (cases of ataxic hemiparesis, with contralateral lesions of the basis pontis involving leg corticospinal and ann pontocerebellar fibers, according to Fisher, 1978). This is not to say that the cerebellar cortex, the inferior olive, and the mossy fibers are equivalent or equipotential in their control of learning, but only that they are all necessary. Their roles are quite different, but the differences are only to be revealed by integrated modeling and experiments that asks questions about each. 2. MICROCOMPLEXES AND THE MODULATORY ROLE OF THE CEREBELLUM In modeling the cerebellum, we stress that the cerebellar cortex and nuclei form an integrated system, and we view this system as divided into small structural and functional units inserted into various extracerebellar systems. Figure 3 reproduces one of these units, named a cerebellar corticonuclear microcomplex (Ito, 1984). A microcomplex is composed of a cerebellar microzone and a small number of nuclear cells and receives (to simplify) two kinds of input, mossy fibers and climbing fibers, the output being carried by the deep nuclear cells. Both mossy fibers and climbing fibers supply collaterals to the nuclear cell group as well as passing to the corresponding microzone of cerebellar cortex. The set of mossy fiber inputs are transformed by the granules cells whose axons form the parallel fibers. A typical Purkinje cell may receive input from on the order of 100,000 Microzone (cerebellar cortex) pf gc PC mossy fibers input output error signal Figure 3. A corticonudear microcomplex,the structural-functional unit of the cerebellum, involvinga patch of cerebellar cortex and the patch of cerebellar nucleus to which its Purkinje cells project, cf, climbingfiber; PC, Purkinjecell; gc, granulecell; 10, inferiorolive. (Adaptedfrom Ito, 1984).

18 M.A. Arbib, N. Schweighofer & W.T. Thach parallel fibers yet will always receive input from only one climbing fiber (the axon of a cell in the inferior olive, IO). The parallel fibers are long enough (Mugnaini ~'}83) to provide synapses across many other microzones. The set of parallel fibers crossing a given microzone constitutes a general context for the present sensorimotor actions in the form of a large set of signals providing information about the state of activity of various structures, from the higher level to sensory ones. The granule cells ensure that the parallel fibers each carry some combination of activity on several mossy fibers, rather than simply relaying their activity. We now note several crucial facts and hypotheses: 1) The only output cells of a microzone are Purkinje cells (PCs). As PCs have inhibitory action upon nuclear cells (while collaterals of mossy fibers excite the nuclear cells), the signal flow from the nuclear ceils is modulated by the microzone action. 2) The climbing fibers are commonly considered as error detectors and evidence has been accumulated for this (see Ito (1990) for a review). Thus, climbing fibers convey signals encoding errors in the performance of the system in which a given microcomplex is installed. 3) Climbing fiber signals induce LTD (long-term depression of synaptic strength) in those parallel fiber~PC synapses which were coactivated with the climbing fibers (within a certain time window). In the next section we will see how this general mode of function can be applied to model the saccadic system. The model is due to Schweighofer et al. (in press), to whom we refer the reader for a more detailed review of the experimental data which ground the model. In Section 4 we will vary that model to account for the adaptation of throwing. 3. MODELING THE ROLE OF CEREBELLUM IN SACCADE ADAPTATION 3.1 The Structure of the Model Figure 4a shows the overall structure of the model. Goldberg et al. (1993) found that stimulation of SC (the superior colliculus, known to provide a retinotopic control surface for saccades) produces saccades which are not adapted to target perturbations. This result suggests that the path concerned with target location and involving the cerebellum comes from \"higher up\" than the superior coMculus. It could be from the FEF (frontal eye fields of the cerebral cortex), the posterior parietal cortex, or even the visual cortex or the lateral geniculate nucleus. Schweighofer et al. assume that the cerebro-ponto- vermal side path for saccade adaptation starts from the FEF, going through a pontine nucleus, then through the lobules Vie and/or VII of the vermis to the FOR (fastigial oculomotor regions), ending up in the parapontine reticular formation (PPRF) where the brainstem saccade generators reside (at least for horizontal movements).

- - L (LTD~ eligi - Target~ ;~=1 - - k § SUl~riorCol,iculus J l ..... .9 ~ .... / 89 , thalamus ) Target~ Fadonirvdeecgataitozionen lkJsh(a~d1ha7op6tre1izd7od6netsairled|[ Figure 4. System views of cerebeUar modulation of the saccadic system and in the adaptive. Note that \"delayed feedback\" (which is relayed via the [not shown] I0) is a non-adaptive pathway. (b) Putative mechanisms of adjustment between eye position a information on eye position arriving in intermediate zone lobulus simplex is carried o which control eye, neck, arm and hand muscle synergies. As in the saccade model, the is further \"upstream\", via premotor cortex.

delayedfeedback cerebralgating ibility)1 • visualinput A I I ~ --eeYlnt gaze ,,, cerebralgating Oo delayedfeedback l ii l visualinput ~,,,~ JI )J - L CortexJ - oSdheonutaldt.,eorn ~ throw ~,,~o adaption of the throw to deviating prisms: (a) The saccadic system must be form of visual input, but it is segregated from that which serves as input to the and synergy of the muscles in the trunk and arm involved in throwing. Afferent over parallel fibers to purkinje cells which project to cells in the dentate nucleus e cerebellum provides a correction to the main pathway, but here the correction

20 M.A. Arbib, N. Schweighofer & W.T. Thach From Goldberg et al.'s experiments, we infer the existence of one or several neural maps where adaptation occurs at specific spatial positions, based on the positions of the first target: Indeed, the adaptation is selective to a set of saccades with similar amplitude and direction. To account for the crude \"correction map\" found in the cerebellar cortex, we will keep the major thesis of the D&A model (Dominey and Arbib, 1992) that a functional topography that preserves saccade direction and amplitude is maintained through multiple projections between brain regions until it is finally transformed into a temporal pattern of activity that drives and holds the eyes onto the target. The preserved topography is a map coding for amplitude and direction of an eye movement vector that, when combined with the current eye location, will center the eye on the target. The Schweighofer et al. model adds an adaptive component, postulating (in line with the experimental data) that saccadic gain change for a particular region of space (around the targe0 will be accomplished within the functional topography of the granule cell layer and the Purkinje cell layer. We now present the essential features of the model, but refer the reader to the original paper for the equations and parameter settings which constitute the formal description of the model A retinotopic map is sent via the pontine nucleus to a set of mossy fibers we call mfret. To simplify, we assume that the pontine nucleus cells (from which the retinotopic mossy fibers arise) are mere relays in which the precise target map is somewhat lost by divergence-convergence. The spread of activity is modeled by a gaussian distribution of weights to form a \"blurry~ topographic connection from the motor layer of the FEF. This \"coarse coding\" speeds up adaptation (Albus, 1981), allowing one to update a group of cells which are \"close\" to the selected cell so that learning is thus extended to saccades of similar amplitude and direction, as seen in Goldberg et al.'s experiments. In the model, another set of mossy fibers carries eye position. The position coding cells receive proprioception signals from the oculomotor muscles. Another factor to be taken into account is the correlation observed between FOR firing and the saccade duration as well as the anatomical connections from the brainstem SGs (saccade generators) back to the vermis (Yamada and Noda, 1987). These fibers carry a temporal signal, such that the firing of the FOR neurons (as well as PCs) will be somewhat synchronized with the SG's activities. The model represents three types of cerebellar neurons - granule cells, Purkinje cells and FOR (cerebellar nucleus) cells - but no cerebellar interneurons are taken into consideration. The granule cells generate a statistical distribution of combinations of mossy fibers carrying retinotopic signals, two position signals and a temporal signal. Each Purkinje cell receives inputs from all the parallel fibers. The parallel fiber - Purkinje cell weights are modifiable. [For the present study we make the following simplifying

Modeling the Cerebellum: From Adaptation to Coordination 21 assumption: As the climbing fibers, on their way to the cerebellar cortex, send collaterals to the deep nuclei, the excitation of the fastigial neurons by these collaterals will nullify the strong inhibition caused by the complex spikes, which are the response of the PCs to the climbing fiber f'u'ing. The present model thus omits the \"real time\", as distinct from the \"training\", role of the climbing fibers.] The PC axons then converge to the FOR, whose cells they inhibit. Collaterals of the mossy fibers also converge to nuclear cells, and give an excitatory projection. It is the output of the nuclear cells that provides the correction signal to the saccadic generators. We must demonstrate that it can be adapted in the fight direction (i.e., corresponding to the corrective saccade) so long as the initial saccade requires correction. We assume that the post-saccadic information available to correct an erroneous motor command is an encoding of the motor activity (or corollary discharge) for a visually guided corrective saccade and thus posit that the IO receives both sensory and motor information: A pre-IO neuron receives visual and motor inputs and has the role of an \"error detector\". An \"error\" will be detected if a) a target is on the retina but b) not on fovea, and c) a saccade has just been completed. If these three conditions are fulfilled, the output of this neuron \"ungates\" the saccadic IO cells. The error detector system comprises three neurons: a \"goal\" neuron, a \"memory\" cell and a \"pre IO\" neuron. The motivation for having a \"goal\" neuron is the need for a signal that continues until the target acquisition is complete. Thus, this cell starts fuing as soon as the target is not on the fovea and fires until the eye acquires the target: It encodes the goal of the saccadic system. Moreover, when a saccade is generated, some SG signals are sent to a \"movement memory\" cell. These two neurons project to the \"pre IO\" neuron. When active, this neuron \"ungates\" the path from the sensory inputs to the IO. Therefore the corrective saccade will send an error signal to the appropriate microcomplex (direction of the error) with some amplitude information (the IO f'u~.s with a higher probability for a large saccade). 3.2 Adaptation: Problems and Solutions The error information is delayed relative to the efferent signal because the error can only be assessed after the movement has been completed. If, after the first saccade, the absence of a target on the fovea and the presence of a target \"nearby\" signals an error, then adaptation changes the \"gain\" of the agonist-antagonist pair of muscles responsible for the first (erroneous) saccade. This implies the need for a short term memory system capable of retaining the appropriate parameters of the first saccade (position on the retina and corresponding eye position). Moreover, the climbing fiber error signal corresponding to the erroneous first saccade reaches the cerebellar cortex after the signal carried by the

22 M.A. Arbib, N. Schweighofer & W.T. Thach parallel fibers corresponding to the second saccade (see below). The problem is to associate the learning not with the second saccade but with the first one. To address these temporal problems, we will assume synapse eligibility (Klopf, 1982; Sutton and Barto, 1981; Houk et al., 1990). The form postulated by Houk et al. is that activation of a dendritic spine by a parallel fiber leads to the release of a chemical called a second messenger in the spine, where its concentration acts as a short term memory. We say the synapse is eligible if the concentration is above some threshold. If an error signal is provided to the whole cell by the climbing fiber, the resulting increase in Ca++ in the cell will, it is posited, only affect the eligible synapses, and therefore only their efficacies are changed. If a parallel fiber-PC synapse participates in synaptic transmission, it becomes eligible to be weakened by LTD if a climbing fiber signal is sent somewhat later. However, we now add to this view of eligibility the requirement that, when the error signal arrives, the concentration of the messenger should tend to be largest for synapses involved in the initial saccade. We therefore introduce the concept of a \"time window of eligibility\". The example we chose in our model is a concentration over time having the response of a second order system for the concentration [2nd] of the messenger (although, alternatively, we can imagine a second messenger following a first order equation but with a significant delay). Ideally, the concentration matches in time the occurrence of the error signal and the concentration decays relatively fast to ensure a minimum of interference with the next saccade. In the model there are two 1000 elements weight vectors wltd, one for each PC, as each parallel fiber makes a synaptic contact with each PC. With the assumption that the rise in calcium concentration is rapid compared to the second messenger dynamics, the weight update rule is, at each time step, for the ith synapse and for each PC, d_d__wltdi = - a I 0 12ndli dt Here, a is the learning coefficient, I 0 the binary climbing fiber error signal, and wmax > wltdi30. But if this equation alone were operative, all the weights would tend to zero. Thus, we implemented a weight normalization which can be thought of as providing nonspecific LTP (long-term potentiation of synaptic strength), in order to keep the sum of synaptic weights for each PC constant. Cl~e normalization adopted is a subtractive normalization. In simulations we also tried a multiplicative normalization; however, this gave a learning curve somewhat different from the data.) Each weight receives the additional increment: ~i Wltdi n

Modeling the Cerebellum: From Adaptation to Coordination 23 where the sum is over all the weights of one PC, and n = 1000, the total number of synapses per PC. This potentiation has no direct functional role as far as the behavior is concerned, except that it somewhat degrades the performance of the whole system. This effect is reduced by the large number of synapses. In the non-adaptive pathway, a light in the upper half of the retina does not elicit a downward movement since the corresponding signals are not transmitted to the \"downward\" generator. However, to correct for errors, this constrained access is not possible for the adaptive system. For instance, suppose that the first target fails on the upper sector of the map and the second on the lower half. In this case, the first saccade amplitude should be decreased as it is hypermetric. To compensate for this error, a decrease of the agonist innervation pulse and an increase of the antagonist pulse is needed. Therefore, the adaptive saccadic system requires more than simple gain control, but adaptation of coordination between the different saccade generators. Consequently, and as seen in the microstimulation experiments, each microcomplex should be able to influence the SGs of the antagonist and agonist muscle. In other words, for each direction for the first saccade there are two degrees of freedom: Either adaptation occurs in the same direction or in the opposite direction. Adaptation requires coordination. We refer the reader to Schweighofer et al. (in press) for further specification of this coordination, and for the results given by this model. 4. MODELING THE ROLE OF CEREBELLUM IN ADAPTATION OF DART THROWING As already exemplified in our model of saccade adaptation, our theory of cerebeUar function posits that the cerebellar output affects motor repertoires resident in movement generators located elsewhere, and that this effect is not only modulatory, controlling the gain of these MPGs, but also combinatorial - mixing motor elements within and across generators such as to adapt old and develop new synergies of multiple body parts. Coordination is effected by the parallel fibers which via Purkinje cell beams spans the width of up to two different body representations within the deep cerebellar nuclei. We concur with the evidence that the parallel fiber-Purkinje cell synapse is adjustable, under the influence of the climbing fiber, and that this is the probable mechanism for changing synergic combinations. In our model of the adaptation of throwing to wearing prisms, the essential adjustment is proposed to be between eye position and synergy of the muscles in the trunk and arm involved in throwing. The target is seen, and the eye foveates and fLxates the center of the target. Afferent information on eye position (not necessarily excluding visual) arrives in the visual and \"face\" tactile receiving areas in intermediate zone lobulus

24 M.A. Arbib, N. Schweighofer & W.T. Thach simplex (cf. Snider and Stowell, 1944; Snider and Eldred, 1952). Information is carried over parallel fibers to Purkinje cells located more laterally in the hemisphere which project to cells in the dentate nucleus that control eye, neck, arm and hand muscle synergies. With repeated throws, adjustments are made in the strength of the parallel fiber input to Purkinje cells (and possibly cortical intemeurons - basket, stellate and Golgi cells - but they are not included in the present model) such that the changing output produced from the Purkinje cells in response to the eye position (and visual) input modulates the throw sufficiently for it to hit the target. 4.1. Notes on Neural Coding We consider the movement to be decomposed into two parts: 1) aiming and then 2) throwing. Only horizontal shift of gaze by the prisms will be considered, and we thus assume that aiming is realized solely by orientation of the shoulder joint in the horizontal plane. Trunk, elbow and wrist rotation are not taken into account in our model. Only the desired position of the shoulder after aiming (but before throwing) is coded, i.e., we do not model the movement itself, but only the horizontal adaptation of the endpoint of aiming. This is supported by the study of Flanders et al. (1992) who showed that the pointing movement to a target is controlled independently in the elevation and horizontal directions. This parcellation may facilitate comparing a target location signal with signals of the limb position so as to yield a motor error signal. If arm position and target location are represented in a common coordinate system (centered at the shoulder), a simple combination between these internal representations may suffice to compute the initial part of the movement. Note that this parcellation representation in arm reaching is earlier, closer to the sensory side of the nervous system than in the saccadic system. Population studies in the premotor and motor cortices show that cells code for all the possible direction of arm movements in 3D space; no separate coding for elevation or azimuth coding by two sub-populations have been found, which might suggest that the two different processing mechanisms would be coded by the same population of neurons. The segregation would therefore be based on functional assemblies and not spatial ones (Bumod and Caminiti, 1992). We model one such assembly, coding for horizontal desired position. We model the arm (Figure 5) with a reference frame centered at the shoulder, with x pointing towards the target, y to the left, and z upwards. In order to reproduce the data, the minimum model of the arm will have three degrees of freedom: vertical and horizontal shoulder rotations (around y and z, respectively) and elbow rotation. We do not model the third degree of freedom at the shoulder joint of the real ann. The rest position is with the ann along - z with the elbow joint maximally extended, while the endpoint of

Modeling the Cerebellum: From Adaptation to Coordination 25 aiming depends on the strategy, i.e., overarm or underarm, as well as the target location (see Figure 5). In both underarm and overarm strategies, the elbow adopts a characteristic angle prior to the throw phase, and we assume it is the shoulder rotation around z that provides the horizontal component of aiming m whose adaptation we study. Arm Configurations rest position under-hand over-hand shoulder 9 elbow x Shoulder orientation before adaptation with 45 ~prisms (top view) X Target I ~x z Shoulder Figure 5. Simplified model of the arm. Note the two degrees of freedom for the shoulder and one for the elbow. Adaptation in the present model is limited to the horizontal shoulder rotation. During adaptation, the cerebellum shifts the population coding of the desired shoulder position after aiming by an amount opposite to the initial deviation. If we follow the \"Georgopoulos view\" (Georgopoulos, Schwartz, and Kettner, 1986) that premotor and motor cortex neurons code a \"population vector\" S which codes shoulder direction, it is tempting to posit that the nuclear output is a signal corresponding to a vector C in world coordinates which is added to the gaze vector G to yield the \"adapted\" shoulder planar vector S = G + C. Since the shoulder vector should point toward the center of the target, adaptation ends when cos(C, S) = - cos(G, S). However, when the glasses are taken off, G points towards the target, and C does not

26 M.A. Arbib, N. Schweighofer & W.T. Thach change until re-adaptation takes place. In the case of a 40~ prism, the shoulder will point at -40~ yet if S = G + C, the new S would be at - 20-. Therefore this first model cannot account for the data. The actual transformation needed is a vectorial rotation and not a sum. If the cerebellum learns how to rotate the desired shoulder position coding by 40~ removal of the lenses will give an error of 40 ~, as in the experiment. The most common neuron considered in neural network modeling forms (some function of) a linear combination of its inputs. However, a rotation is a bilinear combination of its inputs, as the following very simple example shows. Suppose that the neural and world representation of a vector in 2D is the same, i.e., that any vector can be decomposed on an orthonormal basis and that all vectors are unitary. In this case, S, G and C are neural and physical quantities: If c, g and s are the angles between the different vectors and x, the rotation: g ---> s = c + g is given by cos(s) = cos(c) cos(g) - sin(c) sin(g) and sin(s) = sin(c) cos(g) + sin(g) cos(c) which is a bilinear transformation. The mapping from the 4 inputs to the 2 outputs involve two matrices, of 4 weights each (in our particular case, 4 of these values are 0). Burnod et al., (1991) show that for any two vectorial representation a la Georgopoulos, there exist a bilinear transformation corresponding to a rotation. In the first studies on population coding (Georgopoulos et al., 1986) it was first assumed that neurons were coding the direction of the movement. However, it is more likely (Kalaska et al., 1992; Mussa-Ivaldi, 1988; Sanger, 1994) that the recorded cells are encoding arm movement variables (not only direction) related to the shoulder movement during reach, due to a stereotyped coupling of shoulder muscle activity and joint motions to handpath during normal reaching behavior. Consistent with this is the finding by Caminiti et al. (1991) that cortical cells' preferred directions change with initial shoulder angles. It is to be noted that the distribution of preferred direction across cell population is approximately uniform, suggesting that single cells are not confined to coding shoulder motion variables along one of the 3 cardinal degree of freedom of movement of the shoulder joint. Rather the code seems to involve all degrees of freedom but in different ratios for different cells. 4.2 The Structure of the Model A subject throws where she looks. Before the throw, head and trunk are \"towards the target\". This ability, under normal circumstances, does not involve the cerebellum, as is shown by the \"accurate\" throwing (the mean is on target, though the variance is high) by the cerebellar patient when not wearing prisms. Yet, from the inability of cerebellar

Modeling the Cerebellum: From Adaptation to Coordination 27 patients to adapt their throw to prisms (Martin et al., in press), we infer that the shoulder area of the dentate - - which is part of the lateral cerebellar system B is necessary for the aiming adaptation, in accord with the putative role of the dentate in planning and preparation of the movement. Moreover, the cerebellum is involved in a side path projecting to the premotor cortex, area 6, via the ventral thalamus (Shinoda et al., 1992). As shown by the underarm vs. overarm experiments, the adaptation is not a recalibration of the visual coding of space, i.e., the transformation from eye-head coordinates to body-centered coordinates (which would be task neutral) but acts on a direct transformation of coordinates specific to the task. Each motor schema is a controller with its own more or less private coordinate system. Figure 4b gives an overview of the system: The premotor cortex \"prepares\" information on desired horizontal shoulder position for the motor cortex. Since the lateral cerebellum projects (via the dentate nucleus) to premotor cortex, we postulate that it is this cerebellar signal that adjusts the premotor cortex appropriately to changing circumstances (e.g., prisms), and we here model how the cerebellar circuitry can adapt on the basis of a delayed error signal provided by vision of where the dart lands relative to the target. The system (human + dart) is an open loop system since the error in dart throwing is available only after the movement and so the error cannot be used to correct the given movement; however, over trials, the correct match between gaze and throw is learned. Before throwing, the subject foveates the target and therefore the internal representation of gaze, and hence of the desired shoulder position (before it undergoes adaptation), is changed. This information is coded in a distributed manner, providing robustness to lesion and noise: A signal distributed over many noisy nonlinear channels may be summed to yield an accurate signal. This system has been modeled using leaky integrator neurons in our NSL simulation environment. To reproduce all the experiments, the mossy fiber inputs that we consider are (Figure 6): (~) (b) 1 Figure 6. This row of cells shows the mossy fiber inputs. Three different moralities are encoded, each carded by a distinct sub-population of neurons. From right to left, the peaks of mossy fiber activity encode the horizontal shoulderposition (derived from the horizontal gaze direction) in the case (a) of the 25~ prism on and (b) for a case without prisms; the vertical shoulder position which is in overhand position; and the cognitive inputs which corresponds to \"mental set\" (e.g., the knowledge that one is wearing prisms). There are 180 neurons, each f'Lringin the 0 - 60 spike/s range.

28 M.A. Arbib, N. Schweighofer & W.T. Thach 1) Desired arm configuration at the end of aiming. This position is calculated from eye position muscles or from a corollary discharge for the control of gaze. Before aiming, proprioceptive inputs of the end-point are not available. Instead, a desired vertical shoulder position should be available to the cerebellum via mossy fibers (this is coherent with fact that there are no direct inputs from the periphery to the lateral cerebellum). This desired position is well learned and cerebellar patients can throw well if no adaptation is required. 2) Desired vertical shoulder position (to distinguish underarm from overarm throws). 3) Cortical projections for some form of \"mental set\". This does not have to explicitly code knowledge for the present purpose, but must differ depending on whether the subject is or is not wearing prisms, and whether the prisms are known or unknown. This input is necessary to explain the ability of highly practiced subjects to immediately switch \"gain\" when donning or removing known prisms. The cortical input to the cerebellar cortex is known to be large. Indeed, the cortico-pontine fibers form a very large group which arises from the whole cerebral cortex. The data shows that \"prism knowledge\" has a large influence on the response, and this knowledge has an orthogonal representation at the parallel fiber level. Orthogonality of the inputs is a important issue in this model, as will be discussed in conclusion. We have earlier seen that the adaptation should act by coding a rotation of the shoulder vector, rather than providing a correction vector which is to be added on. The model provided here combines the gaze vector G with a rotation signal coming from the ventral thalamus using neurons (postulated to be in premotor cortex) which perform not only additions but also some kind of multiplication. Such neurons have been proposed on theoretical grounds (the sigma pi neurons of Rumelhart et al., 1986); and highly non- linear neurons such as the large pyramidal cells or Purkinje cells are supposed to have dendritic spikes which allow them to function as event detectors (Andersen et al., 1987; Burnod et al., 1991; Gluck et al., in press). It is to be noted that the transformation need not be too accurate and that learning and modulation can compensate for non-perfectly realized mathematical requirements. The algorithm we use to perform this rotation was developed by Hoff (1987, unpublished work) and uses sigrna-pi neurons. Burnod et al., (1991) reached a similar result using cortical columns, whose output are a combination of sums and product of the inputs. We have found that similar computations using Gaussian representations of the variables yields comparable results, and that even spatial \"humps\" of activities yield satisfying results. We next turn to the error detection system. With Flanders et al. (1992), we assume that the target location and arm position share the same coordinate system, with error \"in

Modeling the Cerebellum: From Adaptation to Coordination 29 register\" with the shoulder position. A leftward error activates a \"leftward\" group of IO cells. These cells receive a weighted retinotopic projection from the retina so that a large error will give rise to several spikes. Therefore, the climbing fibers fire to give the direction and amplitude of the error given by a visual projection to the IO which retains some retinotopy. A large error will activate (with a certain probability) different PCs than will a smaller error. However, there is a gradient of cf firing activities within each microzone. In a more complete model, vertical shoulder position should also be able to undergo adaptation (to vertical prisms for instance, or muscle lesions, aging etc.). The same parallel fiber set would therefore overlap four microzones instead of the two in the present model and the IO would encode horizontal as well as vertical errors. 4.3. Adaptation The purely feedforward nature of the movement (the error is not corrected during the movement as the error is not known before the dart hits the wall holding the target) and the delay between motion generation and error detection require again the concept of eligibility Q especially since the throw is made in between the aiming and the receipt of the error signal. In the present case, we use a variant of the eligibility model developed for saccadic adaptation. With wltd the vector of adjustable weights for parallel fiber-PC synapses, the synaptic adjustment rule we consider is: dwltd = {- c.lO + d.(1 - 10) }.[2nd] dt with wmax>wltd > 0, c>>d>0, 10 the binary climbing fiber error signal, c the learning coefficient (LTD) and d the \"forgetting\" coefficient. Some kind of LTP in the learning rule is necessary: If only LTD were occurring, all the weights would tend to zero. The difference from the saccade rule is using the d term to achieve this. As the climbing fiber fh-es only if an error occurs, the weights increase most of the time, but very slowly since c>>d. As a consequence, if the system is tuned at a certain moment, the increase of the weights will sooner or later induce an error. One (or a few) corrective saccades will then be generated, and because c>>d, the weights soon regain the correct values. Each variant of the eligibility-based learning rule has certain advantages, and it is a topic of current research both to explore their properties formally and to define new experiments which can better determine the rule that best describes plasticity of these synapses in the real cerebellum.

Figure 7. These three simulations show the spatial behavior of the cerebellar neuron known prism: Each row graphs the activity of a layer of cells just prior to the throw. First row: The mossy fiber input. Second row: The fETingrates of 15 PCs (vertical scale 0-110 spikes/s, as in the follow Third row: The nuclear cells firing rates. The nuclear cells are inhibited by the PCs, Fourth row: The response of the 20 thalamic neurons. Last row: The 40 premotor neuron activities. The shoulder position is derived from t (a) Situation before the first throw with prism on (corresponding to (a) in figure connections from the parallel fibers with initial random weights. The background fir deviated from the middle due to the gaze input. The premotor activity corresponds to activity peak doesn't represent the shoulder direction, as the latter is given by the \"ce Co) End of adaptation to a prism. The depressed PC activities (on the left) releas distributed activity is pushed back in the middle (the corresponding shoulder position (c) After re-adaptation to the non-prism situation (corresponding to (b) in figure 6). but re-learning that occurs. Also note the quasi-constant thalamic total activity ( thalamic complex, which results in contrast enhancement and consequently in a bette

r~ ns in the course of adaptation to 25 ~ prisms and underhand throwing with a wing groups of neurons). , and are driven by a high background rate. the premotor layer activity by the population vector transformation. e 6). The apparent uniform background rate of the PCs is due to random ring rate in the nucleus provides facilitation to the premotor activity which is o a 25 ~ angle between the shoulder and the forward direction. (Note that the enter of mass\"). se activity in the nuclear cells and in the thalamic neurons. The premotor n is 0~ The PC layer shows another depression on the right side: It is not forgetting (during adaptation) due to the negative feedback achieved by the reticular er shift of the premotor \"hump\". The shoulder angle is 0~ again.

Modeling the Cerebellum: From Adaptation to Coordination 31 4.4. Results Simulated prism adaptation experiments are shown in Figures 7-9. Note that with a model performing a pure rotation, the off-prism initial error is exactly opposite to the initial with- prism error (Figure 7). Both forgetting and relearning are present in the system: forgetting is included in the learning rule and is due to normalization. If there is no climbing fiber activity at a particular site, the weights are very slowly increased. As the error signal carded by the climbing fibers decreased the weights, there is forgetting of the previously learned patterns. By contrast, re-adaptation to the non-prism situation after adaptation to prisms is due to learning rather than forgetting, as can be seen on Figure 7c. In the PC layer, there is a large depression after complete re-adaptation. Figure 8 shows the trial-by-trial behavior corresponding to the neural adaptation shown in Figure 7 for three of the trials. Experience with a -25 degree prism 25 ...... 15 A lO l 'lID s Q | ~.~ O -15 ~ trials Figure 8. Simulated 25~ prism adaptation experiments. This figure and the following show the adaptation of angle between the shoulder angle (derived from the premotor activity) and the \"forward\" direction after aiming over trials. The adaptation requires 20 trims, somewhatmore than re-adaptation to the non-prism condition. Notethat the irregularities in the curve are due to the probabilistic firing of the climbingfibers. Figure 9 show the Over/Underhand experiments. Note that in this case the vertical shoulder position is different for the two throwing strategies. The model reproduces the experimental data reported by Thach: There is some transfer from overhand to underhand in

32 M.A. Arbib, N. Schweighofer & W.T. Thach the prism adaptation.., though in some subjects there is no transfer, while in others the transfer is total. We adopted a middle ground with some overlap in the mossy fiber inputs between the two positions and a not too large mossy fiber input for the vertical shoulder position. Overhand I Underhand o0 J G oS -S \"0=-I0 Q ,'o I , i I 20 30 40 SO trials Figure 9. Over/Underann experiment. In this case, the vertical shoulder position is different for the two throwing strategies. The fu'st part of the learning curve shows adaptation with an overhand strategy. After the prisms are removed, an underhand throw strategy is used. Then 0ast peak) throwing is made overhand again. The model reproduces a typical case reportealby Martin et al: There is only partial transfer from overhand to underhand in the prism adaptation. 5. DISCUSSION The proposed model embeds the cerebellum in a very general framework, applying to both saccade adaptation and dart throwing. Even though the direct mapping from sensory to motor output is somewhat plastic, adaptation to novel context does not occur reliably without the cerebellum. Our model uses the same cerebellar model for two different types of adaptation, with a similar coding of the error n but in the dart throwing model the cerebellum projects \"upstream\" to the premotor cortex instead of \"downstream\" to the brainstem. The microzone concept holds in both cases. The unification of diverse information - from sensory signals to cerebral codes for \"mental set\" - is made possible by the large number of granule cells, each of which forms a sample of diverse mossy fiber signals. Coordination of the modulation of different MPGs is made possible by the long length of the parallel fibers, overlapping different microzones. One sees a tendency for the

Modeling the Cerebellum: From Adaptation to Coordination 33 cerebellum to be less and less concerned with the actual movement when moving from vermis to the \"new\" lateral hemispheres. Finally, we note that the concept of eligibility addresses a key problem in the analysis of adaptive behavior: how can reinforcement or error signals to a network affect those cells which were active some time earlier? Extending the ideas of Klopf, Sutton and Barto, and Houk, we suggest that a short-term memory internal to individual synapses may provide a \"window of eligibility\" when the delay between activity and feedback is on the order of a few hundred milliseconds. Although it takes us beyond the reach of the present study, we note that a more explicit form of short-term memory seems required to link events more widely separated in time. Important clues for future modeling, and for the design of adaptive systems, may come from the phenomenon of trace conditioning. Here, an animal without cerebellum cannot be conditioned in a simple conditioned response; and an animal with cerebellum and without hippocampus can be conditioned only if the delay between unconditioned and conditioned stimulus is at most a few hundred milliseconds. The animal must have both cerebellum and hippocampus intact if it is to be conditioned when this delay is much longer (Moyer, DeYoe, and Disterhoft, 1990). The hypothesis is that the hippocampus holds a trace during the intervening period, bringing yet another neural network into play. ACKNOWLEDGMENTS The research at USC was supported in part by Grant N00014-92-J-4026 from the Office of Naval Research for research on \"Cerebellum and the Adaptive Coordination of Movement\". Appendix: List of Abbreviations cf Climbing fiber D&A Dominey and Arbib model FEF Frontal eye fields FOR Fastigial oculomotor regions LTD Long term depression LTP Long term potentiation SC superior coUiculus PPRF Paramedian pontine reticular formation IO Inferior olive SG Saccade generator PC Purkinje cell

34 M.A. Arbib, N. Schweighofer & W.T. 1bach REFERENCES Albano, J.E., & King, W.M. (1989). Rapid Adaptation of Saccadic Amplitude in Humans and Monkeys. Visual Sciences, 30, 1883-1893. Albus, J. (1981). Brains, Behavior and Robotics. Peterborough, NH: BYTE books. Andersen, R.A., Essick, G.K., & Siegel, R.M. (1987). Neurons of area 7 activated by both visual stimuli and oculomotor behavior. Journal of Neuroscience, 3, 532-548. Baizer, J.S., & Glickstein, M. (1974). Role of the cerebellum in prism adaptation. Journal of Physiology (London), 236, 34-35. Burnod, Y., & Caminiti, R. (1992). Cortical mechanisms of visuomotor transformations underlying arm movements to visual targets. Behavioral and Brain Sciences, 15, 332- 333. Burnod, Y., Grandguillaume, P., Otto, I., Ferraina, S., Johnson, B.P., & Caminiti, R. (1991). Visuomotor transformations underlying arm movements toward visual targets: A neural network model of cerebral cortical operations. Journal of Neuroscience, 4, 1435-1453. Caminiti, R., Johnson, P.B., GaUi, C., Ferraina, S., & Burnod, Y. (1991). Making arm movements within different parts of space: The premotor and motor cortical representation of a coordinate system for reaching to visual targets. Journal of Neuroscience, 5, 1182-1197. Dominey, P. F., & Arbib, M. A. (1992). A Cortico-Subcortical Model for Generation of Spatially Accurate Sequential Saccades. Cerebral Cortex, 2, 153-175. Flanders, M., Helms, Tillery, S.I., & Soechting, J.F. (1992) Early stages in sensorimotor transformation. Behavioral and Brain Sciences 309-362. Georgopoulos, A.P., Schwartz, A.B., & Kettner, R.E. (1986). Neuronal Population Coding of Movement Direction. Science, 1416-1419. Gluck, M.A., Goren, O., Myers, C., & Thompson, R.F. (in press). A higher-order recurrent network model of the cerebellar substrates of response timing in motor- reflex conditioning. Journal Cognitive Neuroscience. Goldberg, M.E., Musil, S.Y., Fitzgibbon, E.J., Smith, M., & Olson, C.R. (1993). The Role of the Cerebellum in the Con~ol of Saccadic Eye Movements. In N. Mano, I. Hamada & M.R. DeLong (Eds.), Role of the Cerebellum and Basal Ganglia in Voluntary Movement (pp. 203-211). Amsterdam: Excerpta Medica. Houk, J.C., Singh, S.P., Fisher, C., & Barto, A.G. (1990). An adaptive sensorimotor network inspired by the anatomy and physiology of the cerebellum. In W.T. Miller, R.S. Sutton, and P.J. Werbos (Eds.), Neural Networksfor Control, The MIT Press. Ito, M. (1984). The Cerebellum and Neuronal Control. New York, Raven Press.

Modeling the Cerebellum:From Adaptation to Coordination 35 Ito, M. (1990). A New Physiological Concept on Cerebellum. Review of Neurology (Paris) 146, 564-569. Kalaska, J.F., Crammond, D.J., Cohen, D.A.D., Prud'homme, M., & Hyde M.L. (1992) Comparison of cell discharge in motor, premotor and parietal cortices. In R. Caminiti, P.B. Johnson, Y. Burnod, (Eds.), Control of arm movement in space. Berlin: Springer-Verlag. Klopf, H. (1982). The Hedonistic Neuron. A theory of memory, learning and intelligence. Washington D.C: Hemisphere. Martin, T., Keating, J.G., Goodkin, H.P., Bastian, A.J., & Thach, W.T. (in press). Prism Adaptation Of Human Eye-Hand Coordination: Task Specificity And Dependency On The Olivo-Cerebellar System. Moyer, J.R.Jr., DeYoe, R.A., & Disterhoft, J.F. (1990). Hippocampectomy disrupts trace eyeblink conditioning in rabbits. Behavioral Neuroscience, 104, 243-252. Mugnaini, E. (1983). The length of cerebellar parallel fibers in chicken and rhesus monkey. Journal of Comparative Neurology, 220, 7-15. Murphy, M.G., & O'Leary, M. (1971) Neurological deficits in cats with lesions of olivocerebellar system. Archives of Neurology, 24, 145. Mussa-Ivaldi, F.A. (1988). Do neurons in the motor cortex encode movement direction? An alternative hypothesis. Neuroscience Letters, 91, 106-111. Noda, H., Sugita, S., & Ikeda, Y. (1990). Afferent and efferent Connections of the Oculomotor Region of the Fastigial Nucleus in the Macaque Monkey. Journal of Comparative Neurology, 302,330-348. Optican, L. (1982). Saccadic dysmetria. In G. Lennerstrand, D. Zee & E. Keller (Eds.), Functional Basis of Ocular mobility disorders (pp. 441-451). Oxford: Pergamon. Ritchie, L. (1976). Effects of cerebellar lesions on saccadic eye movements. Journal & Neurophysiology, 39, 1246-1256. Robinson, D.A. (1987). Why visuomotor systems don't like negative feedback and how they avoid it. In M.A. Arbib & A.R. Hanson (Eds.), Vision, Brain, and Cooperative Computation (pp. 89-107). Cambridge, M.A.: A Bradford Book/MIT Press. Rumelhart, D.E., & McClelland, J.L., (Eds.) (1986) Parallel Distributed Processing - Explorations in the Microstructure of Cognition. Cambridge, MA: The MIT Press. Sanger, T.D. (1994). Theoretical consideration for the analysis of population coding in motor cortex. Neural Computation, 6, 29-37. Schweighofer, N., Arbib, M.A., & Dominey, P.F. A model of the role of the cerebellum in the control of saccades. Submitted for publication.

36 M.A. Arbib, N. Schweighofer & W.T. Thach Shinoda, Y., Futami, T., & Kakei, S. (1992). Inputs from the cerebellar nuclei to the forelimb area of the motor cortex. In R. Caminiti, P.B. Johnson, & Y. Bumod (Eds.), Berlin: Springer-Verlag. Snider, R.S., & Eldred, E. (1952). Cerebro-cerebellar relationships in the monkey. Journal of Neurophysiology, I5, 27-40. Snider, R.S., & Stowell, A. (1944) Receiving areas of the tactile, auditory, and visual systems in the cerebellum, Journal of Neurophysiology, 7, 331-357. Strata, P. (1987). Inferior olive and motor control. In M. Glickstein, et al., (Eds.), Cerebellum and Neuronal Plasticity, (pp. 209-224). New York: Plenum Press. Sutton, R.S., & Barto, A.G. (1981). Toward a modem theory of adaptive networks: Expectation and prediction. Psychological Review, 88, 135-170. Weiner, i . J . , HaUet, M., & Funkenstein, H.H. (1983). Adaptation to lateral displacement of vision in patients with lesions of the central nervous system. Neurology, 33, Yamada, J., Noda, H. (1987). Afferent and Efferent Connections of the Oculomotor Cerebellar Vermis in the Macaque Monkey, Journal of Comparative Neurology, 265, (pp. 224-241).

Motor Control and Sensory Motor Integration: Issues and Directions 37 D.J. Glencross and J.P. Piek (Editors) 9 1995 Elsevier Science B.V. All fights reserved. Chapter 3 INTERACTION OF THE BASAL GANGLIA AND SUPPLEMENTARY MOTOR AREA IN THE ELABORATION OF MOVEMENT R. lansek, Geriatric Research Unit, Kingston Centre, Cheltenham, Victoria J.L. Bradshaw, J.G. Phillips, R. Cunnington Psychology Department, Monash University M.E. Morris Geriatric Research Unit, Kingston Centre, and Schools of Physiotherapy and Behavioural Health Sciences, La Trobe University This chapter reviews basal ganglia (BG) function and concentrates on the BG and supplementary motor area (SMA) interaction. A framework for this interaction is described based on four general areas of research: extrinsic anatomical connections of the BG; cerebral blood flow studies of the motor cortical regions during movement performance in human subjects; single cell recordings from the BG and SMA in animals and studies of movement performance in subjects with Parkinson's disease. This review suggests that the motor function of the BG is twofold and that both of these functions are expressed via the motor cortical regions. Firstly the BG provide internal motor cues that enable the release of submovements from the SMA for execution by the motor cortex. The cue (phasic neuronal activity) interacts with the SMA (sustained neuronal activity) to string submovements together in the correct timing sequence. The second function is to contribute to cortical motor set (sustained neuronal activity) which maintains whole movement sequences in readiness for running and execution. This contribution may be to the SMA, premotor area or to both. The BG is only utilized in these two functions when the movements or sequences are skilled and require few attentional resources for their performance. In Parkinson's disease a defective cue leads to slowing of skilled movement sequences and associated instability of submovements (each submovement cumulatively decreases in amplitude and velocity). This is the phenomenon of hypokinesia. A defect in the contribution to motor set leads to an inability to initiate whole skilled movement sequences (akinesia). 1. INTRODUCTION The motor function of the basal ganglia has eluded researchers for most of this century. Wilson (1912) first documented the involvement of the basal ganglia nuclei in the aetiology of movement disturbance when he described hepato-lenticular degeneration, a progressive disease that causes cirrhosis of the liver and degeneration of the basal ganglia leading to muscle rigidity and tremor. On reviewing the function of the basal

38 R. lansek et al. ganglia he alluded to the lack of knowledge of its role in movement by referring to the basal ganglia nuclei as demonstrating the characteristics of basements, namely darkness. Although clinical neurology has provided numerous examples of motor disturbance through disease involving the b~al ganglia, the role of these nuclei in movement control has not yet been fully elucidated. To some extent, Wilson's original comments still hold true. The basal ganglia are a collection of deep nuclei located within the brain and brainstem. They include the caudate nucleus and putamen (together constituting the striatum), globus pallidus, substantia nigra and subthalamic nucleus (Figure 1 in Morris et al., this volume). Various disease states produce movement disorders which could be attributed to pathology of selected nuclear structures. In basal ganglia disease the motor disturbance either involves diminished movement or excessive movements (Wichmann & DeLong, 1993). Associated phenomenon which are seen with either of these two types of movement disorders include tremor and rigidity. Generally lesions of the caudate nucleus, such as in Huntington's disease, result in chorea (flick like movements) (Wichmann & DeLong, 1993). Lesions of the subthalamic nucleus produce violent choreiform movements involving all body parts on the opposite extremity (Carpenter & Peter, 1972). In contrast, lesions of the substantia nigra can lead to bradykinesia, which refers to slowness in executing movement which is associated with an increase in movement time (Phillips, Bradshaw, Iansek & Chiu, 1993). In addition they can lead to akinesia, a slowness in initiating movement associated with increased reaction time (Phillips et al., 1993). Bilateral pallidal lesions can also result in marked akinesia (Richter, 1945). Dystonia has been less specifically localized to any component structure, usually failing to reveal any morphological abnormality. Yet despite these clinical correlations, we still have little direct knowledge of the normal functions of the basal ganglia and even less knowledge of the pathophysiology of movement disturbance seen in disease states which involve these deep nuclei. Over the last 20 years, however, with advancing knowledge in brain structure and function, a framework has evolved which provides some understanding of basal ganglia function in health and disease. The purpose of this chapter is to present this model and to outline the theory on which it is based. Findings from five main areas of research converge to provide evidence in support of the model. These include the extrinsic anatomical connections of the basal ganglia; cerebral blood flow studies in human

The BG and SMA in the Elaboration of Movement 39 subjects trained to perform specific movement tasks; single cell recordings from animals; investigations using movement potentials of the brain and studies in the experimental psychology of movement performance in Parkinson's disease. A key focus of this chapter is the interaction between the basal ganglia and the supplementary motor area (SMA) in the elaboration of movement and how this interaction occurs at a neural level. We also discuss how disturbance to this interaction can lead to movement disturbance in Parkinson's disease. 2. ANATOMICAL CONNECTIONS A vast amount of information exists regarding anatomical connections of the basal ganglia. To some degree the complexity of these connections could be seen to mask any unifying concept of function rather than to clarify the motor control mechanisms specific to this region. This could be considered to be particularly true of the intrinsic connections of the basal ganglia, where a myriad of reciprocal connections exist between the various nuclear structures (Alexander & Crutcher, 1990; DeLong, 1990; Hazrati, Parent, Mitchell & Haber, 1990). In an attempt to develop a plausible hypothesis of basal ganglia function we therefore consider the intrinsic connections as a unit and concentrate mainly on the extrinsic connections of the major nuclei. Rather than elaborating how motor functions are performed within the unit, the afferent and efferent connections of the basal ganglia will be highlighted as these are the important anatomical features which give major clues on function. The input nucleus of the basal ganglia is the corpus stdatum which is composed of the caudate nucleus and the putamen (Parent & Hazrati, 1993). To a lesser degree, the subthalamic nucleus (STN) receives afferent input from the motor cortex (Hartmann-von Monakow, Akert & Kunzle, 1978). The Output nuclei are both the internal segment of the globus pallidus (GP) and the substantia nigra pars reticulata (SNpr) (Conde, 1992; Parent & Hazrati, 1993; Alexander & Crutcher, 1990). The inputs to the striatum tend to be organized into motor and non-motor components. The motor component involves predominantly the putamen and the non-motor component mainly involves the caudate nucleus. The input to the putamen comes from motor cortical regions which include the motor sensory cortex (MSC), the supplementary motor area (SMA) and the premotor area (PMA) (Yoshida, Nambu & Jinnai, 1993). These projections are somatotopically

40 R. lansek et al. organized so that the leg area is represented in the dorsal component of the putamen, the face is represented in the ventral component, and the ann area in-between. The caudate nucleus also receives inputs from non-motor areas of the cerebral cortex in a topographical manner. The frontal region projects to the head of the caudate nucleus, the parietal cortex to the body, and the temporal cortex to the tail. Reciprocally connected cortical areas tend to interdigitate in the termination field within the caudate nucleus (Kemp & Powell, 1970; Yeterian & Van Hoesen, 1978). A second afferent projection derives from the intralaminar nuclei of the thalamus. This projection is also subdivided into motor and non-motor components. In the monkey the nucleus centrum medianum projects in a somatotopic manner to the putamen with a collateral projection to the motor cortex (Sadikot, Parent, Smith & Bolam 1992). Both projections terminate in the same somatotopic areas, so that the projection to the ann area in the putamen has a collateral projection to the arm area of the motor cortex and the motor cortical arm area projects to the arm area of the putamen. The nucleus parafascicularis of the thalamus projects to the caudate nucleus, in a topographic manner similar to the cortical projection (Jones, Coulter, Burton & Porter, 1977). The motor cortex also projects in somatotopic fashion to the subthalamic nucleus (Matsumura, Kojima, Gardiner & Hikosaka, 1992). The afferent connections of the basal ganglia thus segregate the structure into motor and non-motor components, and some form of somatotopic arrangement is maintained within the basal ganglia as a result of these connections. The output of the basal ganglia derive principally from the internal segment of the globus pallidus (GPi) and the substantia nigra pars reticulata (SNpr) (Conde, 1992; Schell & Strick, 1984; Parent & Hazrati, 1993). The intemal segment of the globus pallidus projects to the ventrolateral thalamic tier (Tokuno, Kimura & Tanji, 1992). The major projection in the monkey is to VLo (Figure 1, Morris et al., this volume), with other projections including VLc and VApc (Schell & Strick, 1984). The SNpr projects to the ventroantedor (VA) nucleus of the thalamus, pars magnocellularis (VAmc) and ventrolateral (VL) nucleus pars medialis (VLm) (Carpenter & Peter, 1972). No overlap occurs in the thalamus between paUidal and nigral inputs (Schell & Strick, 1984). Brainstem projections also occur from the GP and SNpr to the pedunculo-pontine nucleus and the colliculi to the SNpr (Kim, Nakano, Jayaraman & Carpenter, 1976). Projections to the habenular and the intralaminar nuclei of the thalamus also exist, but

The B G and SMA in the Elaboration of Movement 41 these latter projections are minor in size comparexl to the thalamic projections (Kim et al., 1976). It is still unclear whether segregation occurs in the pallidum between motor and non-motor components of the striatum, and whether any segregation is translated to the thalamic projection nuclei. Thalamic projections from VL and VA nuclei are to motor cortical regions (Hoover & Strick, 1993); however, there has been dispute as to the degree of segregation at both the thalamic level and the cortical projection areas for cerebellar and basal ganglia inputs (Tokuno et al., 1992). Suffice to say that pallidal output via the VLo projects predominantly to the SMA, MSC and PMA (Hoover & Strick, 1993). The VApc projects to the PMA, while cerebellar inputs are directed to the MSC and PMA via VPLo and nucleus X. There appears to be a projection to the motor cortex and a projection from the cerebellar relay nuclei to the rostral SMA as well as the motor cortex (Weissendanger & Weissendanger, 1985; Hoover & Strick, 1993). Overall, the major output from globus pallidus intemus (GPi) is to the SMA via VLo, and to the PMA via the VA nuclei (Wichmann and DeLong, 1993). The SNpr output via the thalamus is thought to project predominantly to the PMA. It is apparent from the extrinsic connections that the motor component of the basal ganglia influences movement via the SMA and PMA. In order to elucidate the role of the basal ganglia in movement performance we next examine the role these cortical areas play in normal movement performance from studies of cerebral blood flow. 3. CEREBRAL BLOOD FLOW STUDIES The function of the motor cortical areas can be indicated by neuroimaging of normal human subjects. The findings of Roland and colleagues (1980a, b; 1982) have been illuminating in this regard, and have been subsequently confirmed by more recent studies. For these reasons, the findings of Roland will be considered in detail, as they provide a broad picture of the function of these motor areas. Roland et al (1980a, b) trained subjects to perform a number of f'mger tapping movements of increasing complexity and whole ann movement sequences. Movements consisted of repeated tapping of index finger to thumb, constant sustained pincer grip between index finger and thumb, sequential tapping of thumb to index f'mger, middle f'mger, ring finger and little finger followed by the reverse sequence, and whole ann aiming movements using

42 R. lansek et al. the index finger as the pointer. The pointing task was dependent upon external instruction and involved a previously learnt grid based sequence. In the initial studies, imagined movement was also incorporated into the testing procedure. Subjects imagined but did not perform the finger tapping sequences. These initial studies demonstrated that for isolated index finger and thumb movements, both for tapping movements and isometric force, only the motor sensory cortex showed an increase in cerebral blood flow requirements. Supplementary motor area activation occurred in addition to the MSC for sequential finger tapping movements. Imagined finger tapping produced activation of only the SMA and not the MSC since no movement took place. Whole arm pointing sequences produced activation of the PMA as well as the SMA and the MSC. These findings suggested that the PMA was concerned with the selection of motor plans according to motor requirements. The SMA was concerned with running the movement sequence involved in the motor plan, and the MSC executed each component of the movement plan which contributed to the sequence. A subsequent study conducted by Roland et al. (1982) using positron emission tomography (PET) examined cerebral blood flow changes in the basal ganglia for the f'mger tapping sequence. This study found that blood flow increased in the putamen, globus pallidus, and ventrolateral thalamus and confirmed that the motor component of the basal ganglia was involved with the SMA in the running of learnt movement sequences. This coupling between basal ganglia and SMA in the running of movement sequences has been further confirmed by demonstrating an increase in cerebral blood flow in both regions in Parkinsonian patients when hypokinesia was reversed with an injectable direct dopamine agonist drug, apomorphine (Jenkins et al., 1992). The type of movement paradigm used in functional neuro-imaging studies is of utmost importance in establishing the function of the motor cortical areas. Greater activation of the SMA is observed for complex sequential movements which are well learnt, compared with simple repetitive movements (Roland et al., 1980b; Shibasaki et al., 1993), and for internally determined movements compared with externally cued movements (Deiber et al., 1991). Activation of the SMA and basal ganglia also increases as complex sequential tasks become more practised. Seitz and Roland (1992) examined cerebral blood flow changes accompanying learning of a long and complex finger tapping task. They found that during the learning process little activation

The BG and SMA in the Elaboration of Movement 43 occurred in the SMA and the basal ganglia. Most of the activation occurred in diffuse cortical areas, particularly Broca's area, presumably because subjects internally counted the sequence during the learning process. Once the sequence was learnt however, basal ganglia blood flow was increased. These fmdings suggested that the performance of movements utilized different cortical areas, depending upon the novel nature of the task and, by implication, the predictability of the task. Basal ganglia activation occurred only for well learnt movement sequences. These studies therefore suggest that the motor component of the basal ganglia is intimately related to the SMA in the running of movement sequences which are well learnt and predictable. Functional imaging studies, however, failed to reveal the nature of this interaction. Single cell studies in animals, involving recordings made from neurones of both the SMA and the basal ganglia, will next be considered, as these studies provide further information on the nature of the interaction between these two motor systems. 4. SINGLE CELL STUDIES There have been numerous studies of single cell recordings from neurones in the SMA (for example, Tanji & Kurata, 1985; Romo & Schultz, 1992). The discharge properties are quite varied and certainly overlap with those found in the PMA and the MSC. However, one common finding in the SMA is that of premovement activity. This type of activity has been consistently reported and occurs in up to 60% of neurones examined (Tanji & Kurata, 1985). This activity is characterized by sustained neural discharge which builds up in anticipation of an upcoming movement (Figure 1). Once the cue to move is given the discharge ceases abruptly (Figure 1). This discharge has been termed \"set\" related activity as it appears to be related to the preparedness for an upcoming movement. Mushiake, Inase and Tanji (1990) found that set related activity occurs prior to each submovement of a sequence only if the sequence is predictable to the animal. Further, SMA activity was greater for intemally guided movements, compared with externally guided movements (Mushiake et al., 1991). The externally guided movements appeared to be more related to activity in the PMA (Mushiake et al., 1991). Mushiake et al. (1991) found that phasic activity which occurs in time with movement cues, is common in an area anterior to the SMA, which they termed the pre-supplementary

44 R. lansek et al. motor area. Further, Tanji and Kurata (1985) found neurones in which premovement activity occurs before a sequence of movements but not prior to each of the submovements in the sequence. However these neurones were few, compared with those which showed premovement activity prior to each submovement. It is apparent that neural discharge in the SMA is varied, but that preparatory activity predominates and manifests in isolated movements, for each submovement in a sequence, and prior to initiation of a whole movement sequence. In each case the movements and sequences have to be predictable. In addition phasic activity associated with external cues for movement is also evident. SUPPLEMENTARY I I MOTOR AREA , I I I BASAL GANGLIA I I I I MOTOR CORTEX I I MOVEMENT I ,I I I i I I I I I , ' Movement I onset Movement offset Figure 1. Schematic representation of proposed neuronal activity between cortical regions and the basal ganglia for one submovement in a sequence. Single cell activity in the basal ganglia has been documented in numerous reports (eg., Chevalier & Deniau, 1990; DeLong, 1971; Kimura, Aosaki, Hu & Ishida, 1992; Manetto & Lidsky, 1989; Mink & Thatch, 1987); however only those reports which

The BG and SMA in the Elaboration of Movement 45 describe neural discharge in the GP will be considered in this review. This is because the GP is the principle output nucleus of the basal ganglia and neural behavior here would best express overall function rather than the contributory mechanisms to function which may be recorded in other nuclei such as the striatum or subthalamic nucleus. Early studies suggested that neural activity in the GP was coding for basic parameters of movement (eg., Georgopoulos, Delong & Crutcher, 1983). More recent studies have not however confirmed this suggestion. The findings of Mink and Thach (1987) and of Brotchie et al. (199 la, b) provide evidence against this concept and suggested a possible role of the GP in higher aspects of motor control. Brotchie et al. (1991a) found that neural discharge in the GP in relation to movement, was phasic in nature and that sustained premovement and sustained post movement activity, although present, was rare. Phasic activity was found to occur after movement onset and prior to the initiation of the next component of movement in a movement sequence. They demonstrated that this phasic activity temporally correlated with the end of a movement performed in a series of movements. They further demonstrated that the phasic activity was very plastic and developed over several repetitions of the same movement once a new movement sequence was initiated, and was only present when movements were predictable, easy and required few attentional resources. Sustained neural activity was also dependent upon the predictability of upcoming movements (Brotchie et al., 1991c). Brotchie et al. (199 lb) suggested that the basal ganglia and the SMA interacted at a neural level in the following way. A well learnt movement sequence would be initiated in the cortex via either the PMA, SMA or both. Once initiated, the movement sequence would run to completion automatically, using the basal ganglia-SMA interaction. Once the initial submovement was executed, phasic activity would be produced in the basal ganglia (see Figure 2). This phasic activity (cue) to the SMA would terminate set related activity for the next submovement and enable set related activity for the next (third) submovement to be initiated. The execution of the second submovement would produce phasic activity in the basal ganglia which would in turn terminate set related activity for the third submovement. A domino effect would ensue and the sequence would run automatically to completion. Brotchie et al. (1991c) examined this interaction with neural network modelling and demonstrated that both phasic and (tonic) set related activity had to exist in the basal ganglia in order for movement sequences to run automatically.