a hard scienti½c quest: understanding voluntary movementsweb.mit.edu/ajemian/www/daedalus_2015.pdfa...

A Hard Scienti½c Quest:Understanding Voluntary Movements

Emilio Bizzi & Robert Ajemian

EMILIO BIZZI, a Fellow of theAmer ican Academy since 1980 andPresident of the Academy from2006 to 2009, is Institute Professorin the Department of Brain andCog nitive Sciences and Investiga-tor at the McGovern Insti tute forBrain Research at the Massachu-setts Institute of Technology.

ROBERT AJEMIAN is a ResearchScientist at the McGovern Institutefor Brain Research at the Massa-chusetts Institute of Technology.

(*See endnotes for complete contributorbiographies.)

Scientists and nonscientists alike rarely stop to con -sider what is going on in their brains when they per-form a voluntary movement such as reaching for anobject, throwing a ball, or driving a car. Why? Pre-sumably they may realize that translating some-thing as evanescent as a wish to move into musclecontractions must be an awfully complicated pro -cess. Indeed, they are right: the neural processes thatsubserve even the simplest everyday actions are in -credibly complex and only partially understood. Inthis essay we take up the challenge of explaining whatwe know about this fascinating and complex topic.

Let us begin with the basic fact that, in general, ourmovements–even the simplest actions–are accom -plished through activation of a large number of mus-cles. For example, if you are sitting at your desk typ-ing at your computer and decide to turn to pick upa cup of coffee, you will activate, approximately at thesame time, the eye muscles, the numerous musclesin the neck, and the muscles of the shoulder, arm,forearm, and ½ngers. A simple computation wouldshow that your brain has activated at least thirtymus cles. But note that each muscle is made up ofcells called mus cle ½bers, and that each muscle ½berre ceives a neural input via its own nerve ½ber (seeFigure 1). It follows that the number of elements

83

© 2015 by the American Academy of Arts & Sciencesdoi:10.1162/DAED_a_00324

Abstract: In this article we explore the complexities of what goes on in the brain when one wishes to performeven the simplest everyday movements. In doing so, we describe experiments indicating that the spinalcord interneurons are organized in functional modules and that each module activates a distinct set ofmuscles. Through these modules the central nervous system has found a simple solution to controlling thelarge number of muscle ½bers active even during the execution of the simplest action. We also explore themany different neural signals that contribute to pattern formations, including afferent information fromthe limbs and information of motor memories.

controlled by the neural mo tor system isvery large, even during simple move-ments.

Imagine now what must be taking placein the brain of an athlete in the heat of asoccer match, when practically all the mus-cles of the body must be precisely coordi-nated, with little time for preplanning.Clear ly, the soccer player trying to score agoal has neither the time nor the inclina-tion to explicitly formulate the commandsignals to control the millions and millionsof muscle ½bers in his or her body, andmust instead rely on an effective simplify -ing strategy. How our brains cope with thisinherent complexity remains one of thefun damental questions in motor systemneuroscience.

Years ago a group of neuroscientists, in - cluding one of the authors of this essay, de -cided to investigate this basic question bylaunching a series of exploratory searchesaimed at identifying the way in which thecentral nervous system (cns; the com plexof nerve tissues that control the ac tivitiesof the body, comprising the brain and thespinal cord) controls the multitude of mus -cle ½bers that are activated during move-ments.1 We started by focusing on thespinal cord in lower vertebrates and quick -ly found that a special group of cells calledinterneurons–neurons that transmit im -pulses between other neurons, and that areinterposed between the sensory portionof the spinal cord and its motor output–are the key elements that implement thesimplifying strategy.2

84

Under -standing

VoluntaryMovements

Dædalus, the Journal of the American Academy of Arts & Sciences

Figure 1Muscle Fibers and Axons from Motor Neurons

Axons are the long threadlike part of a nerve cell along which impulses are conducted from the cell body to othercells. Source: Frontiere Della Vita (Rome: Istituto Della Enciclopedia Italiana, 1999).

Axons from CNS

fibersMuscle

Muscle(biceps)

Interneurons are organized in function -al modules, and each module activates aparticular set of muscles in distinct propor -tions. We labeled this entity of patternedmuscle activation a muscle synergy. Thismod ular spinal structure is the centralpiece of a discrete combinatorial systemthat utilizes a ½nite number of discreteele ments (that is, the muscle synergies) toex press a voluntary movement. The com -b inatorial system is controlled by the neu -rons residing in the cortical frontal areas(the cerebral cortex covering the frontallobe). Anatomically, the cortical neuronstransmit impulses to select and combinespinal modules. Following the arrival ofcortical command signals, a cascade ofneu ral events ensues: the activated spinalmodules ½re the motoneurons (nerve cellsforming part of a pathway along whichimpulses pass from spinal cord to a mus-cle) and their motor nerves induce a de -polarization of the muscle ½bers, whichin turn is followed by muscle contractionand movements. Researchers can easilyre cord the electromyographic activity(emg)–the depolarization of muscle ½ -bers–with electrodes placed in or on themuscle surface.

To identify the muscle synergies we useda factorization algorithm that takes as in -put all of the muscle emg data and extractsfrom these data both a set of generativemuscle synergies and a coef½cient of eachsynergy during the composition of a par-ticular motor behavior. The experimentalevidence supporting the idea that the cnsuses muscle synergies as output modulesis illustrated in Figure 2. This illustrationshows that a small number of synergiesexplain a large fraction in the variation ofmuscle patterns. In other words, not asmany individual muscle functions need tobe controlled as one might have initiallythought.

Figure 2 shows the emg records for frogsthat are jumping, walking, and swimming.

The upper (unshaded) section of Figure 2lists the names and the emgs of thirteenleg muscles. The shaded wave areas in thissection of the ½gure represent the recti½ed,½ltered, and integrated emgs recorded dur -ing the execution of a single instance ofjumping, walking, and swimming move-ment. The thick line de½ning the contourof the wave represents the outcome of acom putation that reconstructs the musclepatterns by utilizing the muscle synergiesextracted by the factorization procedure.The lower (shaded) section of Figure 2shows the coef½cients of the ½ve synergiesthat were found through the factorization.The coef½cients are placed in a rectangularbox whose width corresponds with syner-gy duration; their position indicates onsetdelay and the height represents amplitudeof emg. Figure 2 demonstrates two impor -tant points: 1) the same synergies are foundto contribute to different movements(note that synergies W1, W3, and W4 are a constituent of both jumping and walk-ing, but with different coef½cients of acti -va tion of emgs; the synergy W5 is used inboth jumping and swimming); and 2) dif -ferent behaviors may be constructed bylinearly combining the same synergieswith different timing and scaling factors.Recent results from the study of musclepatterns during a variety of movements inhumans, monkeys, and other vertebrateshave shown that combining a small set ofmuscle synergies appears to be a generalstrategy that the cns utilizes for simpli-fying the control of limb movements.

The speci½c factorization algorithm thatwe used to extract the underlying synergiesfrom the overall emg data set is known asthe nonnegative matrix factorization.3Oth er factorization algorithms could havebeen used, such as the popular principalcomponent analysis (pca); but neurosci-entist Matt Tresch and his colleagues haveshown that whatever technique one usesto identify the synergies, the end results are

85

Emilio Bizzi &RobertAjemian

144 (1) Winter 2015

essentially the same.4 This suggests thatthe observed muscle synergies are real, asopposed to an artifact of the data analy-sis. Additional observations corroboratethe independent physiological existenceof muscle synergies as fundamental andirreducible units of motor control that arelinearly combined to generate movement.5

The experimental evidence describedabove indicates that the peripheral sectionsof the motor system operate as a discretecombinatorial system. In a way, then, themotor system is like language, a system inwhich discrete elements and a set of rulesfor combining them can generate a largenumber of meaningful entities that are dis -

tinct from those of their elements. Thus,we may have solved the problem of howthe motor system copes with having tocon trol so many different muscles and mo -tor units during the course of a movement:it does so through intelligent modulariza-tion at the level of the spinal cord.

But having proposed a solution for oneproblem, we are immediately led to an -other: how does the brain ½gure out thecorrect combinations of synergies that arerequired to execute a motor act? Certainly,what is impressive about the motor systemis its capacity to ½nd original motor solu-tions to an in½nite set of ever-changingcir cumstances. This capacity is entirely de -

86

Under -standing

VoluntaryMovements


Figure 2Synergies and Variations of Muscle Patterns

The main muscles of synergy W1 are RI, AD, PE, and GA. The main muscles of synergy W2 are SM, VE, PE, and GA.The main muscles of synergy W3 are RI, SM, and VI. The main muscles of synergy W4 are RA, BI, and IP. The mainmuscles of Synergy W5 are ST and IP. These synergies were extracted by pooling together the emgs of three frogsduring jumping, walking, and swimming movements. Source: Emilio Bizzi, Vincent C. K. Cheung, Andread’Avella, Philippe Saltiel, and Matthew Tresch, “Combining Modules for Movement,” Brain Research Reviews 57(2008): 125–133.

RIADSMVI

VERAPEGASTSABIIPTA

W1W2W3W4W5

syne

rgie

sE

MG

sjumping walking swimming

pendent upon the computations per-formed by neural circuitries of the corti-cal areas of the frontal lobe (the lobes oneach cerebral hemisphere lying immedi-ately behind the forehead). These corticalareas generate signals that combine, select,and activate the spinal mod ules. Under-standing these computations has been themain goal of neuroscientists, neu rologists,and psy chologists in volved in the studyof mo tor control. Some progress towardthis goal has been made, but as we will dis-cuss, many of the hard questions remainunanswered.

Our description of the way in which cor -tical commands generate patterns of activ -ity for activating the spinal cord moduleswill begin by considering the major inputsand outputs of the motor cortical regions.

In each frontal lobe hemisphere, thereare at least four major regions concernedwith generating signals for voluntarymove ments: the dorsal and ventral pre-motors (the cortical areas in front of themotor cortex), the supplementary motorarea, and the primary motor cortex.6These highly interconnected regions re -ceive diverse modalities of information(inputs) from a variety of sources, includ -ing: 1) external sensory information aboutthe state of the world (such as visual, au -ditory, tactile information); 2) internalsensory information about the state of thebody (such as muscle length and tendonforce); 3) the executive attentional systemfor determining behavioral saliency; and4) inputs from major subcortical areas suchas the cerebellum and the basal ganglia(whose roles in motor control are some-what obscure). These signals are conveyedto the motor cortical areas where theyconnect to the dendritic tufts of the largeoutput cells of the cortical layers 5 and 6(see Figure 3).7 There these signals aresomehow integrated into a coherent unitto set up a neuronal depolarization, which

is conveyed via the dendritic tree to thecell body of the big output cells in layer 5,and from there via long pathways to thespinal cord.

There are a variety of output pathwaysmade of axons of cortical layers 5 and 6that connect premotor and primary motorcortical areas with a different class of spi -nal neurons. One of these descending path -ways conveys information about an im -pend ing movement, an observation sug-gesting that it may be part of a cortico-spinal circuit contributing to an earlyshap ing of motor commands. Additionaloutput pathways include: 1) cortico-spinal½bers terminating at the level of theinterneurons, which activate the spinalcord modules and therefore are responsi-ble for the expression of the muscle syn-ergies;8 2) a cortico-motoneuronal path-way from the most caudal sections of theprimary motor cortex,9 though we do notyet know how these two descending setsof ½bers cooperate in the execution of vol -untary movements; 3) an important set of½bers connecting the motor cortex to thebasal ganglia; and 4) a set of ½bers connect -ing the cortex to the cerebellum in therecurrent loop, which create a cerebellarpathway whose complex function is pos-sibly related to reentry circuits that con-tribute to shaping the construction of cor -tical patterns of activity. (The function ofthe cerebellum has long been a source ofde bate possibly relating to the function ofcortical patterns activity; see Figure 4.)

But also critical is that the descendingpath ways are mirrored by ascending setsof ½bers forming numerous reentry cir-cuits. Thus, the intimate ties between thecortex and the periphery are an essentialfeature of the “system” for movement.

There is a vast amount of data indicatingthat the motor cortex plays a central rolein generating motor behavior, but there islack of consensus on how neural process-

87


144 (1) Winter 2015

ing within the cortical areas of the frontallobe contribute to voluntary movements.An approach to interpret cortex neural ac -ti v ity, ½rst introduced by Edward Evartsat the National Institutes of Health, wasbased on recording the activity of singlecor tical neurons in monkeys and then cor -relating their ½ring rate with joint mo tion,force, and limb posture.10 Evarts conclud-ed that the motor cortical neurons likelyencoded the muscular force that is neededfor movement. In the early 1980s, neuro-scientist Apostolos Georgopoulos, usinga modi½ed behavioral setup, showed thatcortical neurons recorded from the pri-mary motor cortex were broadly tuned tothe direction of hand movements.11 Thiscorrelation suggests that the motor cortexencodes high-level parameters of move -ment such as direction in task space, rath -er than low-level parameters such as mus -cle forces. But the story did not end there.In the last few decades, researchers have

implicated the motor cortex in the en -cod ing of a litany of different movementvariables, including hand velocity, handposition, joint angles, joint torques, move -ment sequence information, and move-ment curvature. Further, the responseprop erties appear nonstationary, chang-ing with behavioral contexts and choice oftask.12 So what does this all mean? Whatdoes the motor cortex really do?

To put all these observations in perspec-tive, we should consider the limitations ofmicroelectrode recordings. In acute re -cord ing, one or a handful of neurons arerecorded. In chronic recording, an arrayof electrodes is implanted that can recordfrom roughly one hundred neurons simul -taneously. Yet there are millions of neu-rons in the motor cortices, each one high-ly interconnected to other motor corticalneurons and to the many input sourcesthat project to it. The endeavor is thus lim -

88

Under -standing

VoluntaryMovements


Figure 3Pyramidal Neuron

Source: Frontiere Della Vita (Rome: Istituto Della Enciclopedia Italiana, 1999).

basal dendrites

inputstypically arescatteredover theneuron’sdendritictreecell body

apicaldendrite

is the neuron’s outputbranching axon

(the junction is the synapse)

ited both by problems of undersampling(you are only recording from a tiny frac-tion of a population) and problems of lim -it ed sampling information (usually you donot know in what layer a recorded neuronresides, meaning you do not know wherein the context of this complicated andhighly distributed circuit your neuron ½ts).An analogy: suppose that you know noth-ing about how computers work, but areasked to ½gure it out by sticking a volt me -ter into different regions of a computerwhile it runs various programs, recordingits electric potential. Considering this chal -lenge, perhaps it is not so surprising thatthe manner in which patterns form at the

output layer of the motor cortex to gener-ate movements remains a mystery.

Clearly, to move our understanding ofthe motor cortex forward, new approachesare needed. In the last few years, prompt-ed by the urge to look for new avenues,neurophysiologists and computationalneuroscientists have joined forces in orderto make sense of the neuronal recordingdata and then generate theories and mod-els of the motor cortex. The most notableproposed models, such as optimal feed-back control and recurrent neural network,are attempts at formulating a comprehen-sive motor control theory;13 but becausetheir focus is based on our limited knowl-

89


144 (1) Winter 2015

Figure 4The Cortex, Cerebellum, Brain Stem, and a Section through the Spinal Cord

The motor cortical areas are shown with some of the ½bers connecting it to the basal ganglia and the cerebellum.Two cortical spinal pathways are shown, one a direct pathway from the cortex to the spinal motorneurons andanother connecting the motor cortex to the spinal interneurons. Source: Eric R. Kandel, James H. Schwartz, andThomas M. Jessell, eds., Principles of Neural Science, 4th ed. (New York: McGraw-Hill, 2000).

90

Under -standing

VoluntaryMovements


edge of neurons, the resulting models hadonly a modest impact on the ½eld. In gen-eral, these models failed to consider thecomplex interactions among differentclass es of cortical cells and the role of there cur rent circuits that link the cortex withthe spinal cord, basal ganglia, and cerebel -lum. The cortical neurons’ activity shouldhave been evaluated in a different way,since these cells belong to an ensemble.

Fortunately, thanks to developments inmolecular biology, a panoply of new tech -niques is becoming available. For ex ample,imaging techniques are being developedthat will enable an experimenter to recordfrom thousands of neurons simultaneous-ly, while at the same time monitoring theanatomical changes taking place withinthe circuit at the synaptic level. Differentstrains of viruses carrying channelrhodop -sin into targeted populations of neuronsnow make it possible to activate/inhibitneural circuitry. This is an important de -velopment because neural circuits are in -herently parallel and highly interconnect -ed, meaning that it is dif½cult to under-stand what one part of the circuit is doingin isolation of the rest of the circuit. Suchtechnological breakthroughs, togetherwith the development of mathematicaltools for processing and modeling high-dimensional distributed dynamical sys-tems, may change the playing ½eld in sys-tems neuroscience in the years to come.

Neuronal activity in the motor corticalareas is a complex function of sensory in -puts, regional and local interactions amongcells, and cortical reentry circuits. In ad -dition, recent investigations have estab-lished that one more signi½cant way to ½rethe cortical neurons is just to image anaction without producing an actual move -ment. There is also evidence that shiftingfrom one mental task to another changesthe pattern of brain activation. These re -sults were obtained by monitoring region -

al blood flow either with a positron emis-sion tomograph or fmri (functional mag -netic resonance imaging). It is of interestthat most of the activation was found inthe premotor and supplementary areas, butless so for the primary cortex for both mo -tor imagery as well as movement obser-vation. These ½ndings show that the de lib -erate representations of actions involve ac -tivation similar to those occurring duringvoluntary movements. These observa tionsare important as they expand the scope ofthe motor system beyond the gen eration ofactions. Imaging the brain during a motoraction means to evaluate the consequenceof self-programmed movements before ex -ecution, while also providing ways to rep-resent other people’s ac tions.

As a consequence of the intense explo-ration of the motor imaging underlyingphysiology is the realization that cells inthe premotor and primary motor cortexare active when a subject plans an action.This ½nding has opened the way to re -cord from the human cortex and utilizethe neural signals for prosthetic devices.

Closely linked to the motor imaging ofactions is the brain’s representation ofmotor memories; that is, when we learn askill, how is that skill represented in ourneural circuits? Since most of what wedo is guided by what we have learned,this capacity for motor learning embod-ies a crucial facet of our existence. Imag-ine how dif½cult life would become if everytime we engaged in a routine act, like tyingour shoes, we had to perform it with theskill level of a novice. Instead, as we gothrough life we gain facility and acquireexpertise in the form of motor memories:memories of how to perform skilled mo toracts. Where are these motor memoriesstored and how are they represented?

In the case of computers, we know whereand how information is stored. It is storedin the digital switches of transistors that

are housed and addressed in speci½c loca-tions within the computer (the hard driveor random-access memory) or various ex -ternal digital media (such as a disk). In thecase of human declarative memory–which, loosely, is factual information aboutone’s life (names, places, events, and soon)–we also have a fairly good idea. De -clarative memory is stored in the medialtemporal lobe of the brain (a region ofthe cerebral cortex that is located beneaththe lateral ½ssure on both cerebral hemi-spheres of the mammalian brain) includ-ing the hippocampus (the elongated ridgeson the floor of each lateral ventricle of thebrain, thought to be the center of emotion,memory, and the autonomic nervous sys-tem), the entorhinal cortex (an area of thebrain located in the medial temporal lobethat functions as a hub in a widespread net -work for memory and navigation), andperi rhinal cortex (a region of the medialtemporal lobe of the brain that receiveshigh ly processed sensory information fromall sensory regions, and is generally ac -cepted to be an important region for mem -ory). Motor memory, on the other hand,appears to be broadly distributed across allof the major components of the motor cir -cuit, including the motor cortices, the cere -bellum, and the basal ganglia; and how mo -tor memories are stored still remainsmurky, though we have clues.

One clue has come from force ½eld stud-ies in which monkeys adapted their reach-ing movements to different environmentsand perturbations while the activity of sin -gle neurons in their motor cortices was re -corded.14 As expected, when a monkeylearned to move in a novel context, theactivity patterns of the recorded neuronschanged. This ½nding is generally consis-tent with the synaptic trace theory of mem -ory, which says that memories are em -bodied in patterns of synaptic connections(synapses are the junctions between twonerve cells consisting of a minute gap

across which impulses pass by diffusion ofa neurotransmitter) that change in anexperience-dependent fashion, such thatafter the experience, the circuit is capableof generating a new output. However, itwas unexpectedly found that some of theneurons maintained their altered activitypatterns even when the animal stoppedperforming the new task and returned tothe original task. Further, as the behaviorswitched from one task to another, the pat - tern of activity of the neurons changed inan unpredictable fashion.

In a similar vein, a pair of studies usedthe technique of two-photon microscopyto study anatomical changes in the synap-tic connectivity of the mouse motor cortexduring the learning of new motor skills.15

This remarkable technology allows experi -menters to visualize individual synapticspines, the smallest unit of informationtransmission in the brain, over periods ofweeks or months. As expected, these stud-ies showed that learning is indeed accom-panied by the formation of new synapticspines. However, these studies showedsome thing else that was unexpected, evenshocking: when the animals are not learn -ing anything, the synaptic spines are stillturning over at a high rate. In fact, the rateof turnover is so high in the baseline con-ditions that most of the new spines createdduring the formation of the new memorywill be gone in a matter of months or, atmost, a couple of years. Yet these same mo -tor memories are known to persist for theanimal’s entire life.

These observations lead to a profoundparadox. If we believe that memories aremade of patterns of synaptic connectionssculpted by experience, and if we know,behaviorally, that motor memories last alifetime, then how can we explain the factthat individual synaptic spines are con-stantly turning over and that aggregate syn -aptic strengths are constantly fluctuating?How can the memories outlast their puta -

91


144 (1) Winter 2015

tive constitutive components? This is cur -rently one of the great mysteries in motorneuroscience and, in fact, all of systemsneu roscience, reinforced by the dozens oftwo-photon microscopy studies that havefound that regardless of which region ofthe cortex is examined, the synapses areconstantly turning over. How is the per-manence of memory constructed from theevanescence of synaptic spines?

In an attempt to answer this question, werecently developed a new type of neuralnetwork with the distinguishing featureof synapses that are constantly changingeven when no learning is taking place.16

We showed that under certain conditions–conditions that hold during motor learn-ing–the network can stably learn a vari-ety of skills despite these constant weightchanges. The basic point of the model is toreexamine the notion of what constitutesa memory. Neural circuits are highly re -dun dant in that there are many more syn -apses than there are neurons, with eachneu ron being contacted, on average, by tenthousand synaptic spines. Thus, within aneural circuit, many different con½gura-tions of synapses can give rise to the sameinput-output processing. In other words,a network can perform the same functioneven if its synapses undergo change. Withthis in mind, we suggest that a memory,instead of being composed by a ½xed pat-tern of synaptic weights, is actually em -bodied by a ½xed pattern of input/outputprocessing at the level of neural activities.This flexibility gives the system some slackto accommodate synaptic turnover, an in -evitable fact of cell biology, since synaps-es are made of proteins, which have shortlifetimes. Models, like this one, that relyon the stochastic dynamics of complexsystems may prove to be fertile territoryfor understanding recent and future dataon synaptic dynamics.

A ½nal question is whether models likeours can simultaneously shed light on both

the small-scale physiology/anatomy andthe larger-scale behavior of a subject. Af -ter all, the ultimate goal of neuroscience isto mechanistically link the physical en tityof the brain to the more ethereal phenom -ena of the mind. This is not always easy todo and often a model is constructed to ex -plain data in one domain or the oth er, butnot both. Here we have found that a mod -el with perpetually fluctuating synapticconnections may explain an in teresting re -sult from the kinesiology and sports sci-ence community that has been known forover one hundred years.

The ½nding is called the warm-up decre -ment and it can be illustrated as follows: tofunction at a peak performance level, a pro -fessional athlete trained in a ½ne motorskill must practice or warm up for an ex -tended period of time immediately priorto performing. For example, professionalgolfers and professional tennis players willpractice for an hour or more before play-ing in a major competition. In a certainsense, this seems strange. These athleteshave spent much of their lives practicinga particular skill, so why do they still needto practice for so long before competing?A robot that performs a skill needs only tobe turned on and shortly thereafter it willexecute the skill to the best of its abilities.So why do human experts need addition-al practice right before performance?

One possibility is that the practice isneeded to warm up the athlete’s muscles,lig aments, and tendons. However, this the -ory has been refuted by experiments inwhich the body is warmed up by othermeans, resulting in sub-peak performance.Further, athletes of all calibers widely ac -cept that the warm-up ought to use thesame skill set that will be used in perfor -mance. If a professional tennis player wereto practice squash an hour before playinga tennis match, the results would be di -sastrous, though many of the same musclesare used in both activities. So what, then,

92

Under -standing

VoluntaryMovements


93


144 (1) Winter 2015

is going on during this period of warm-up? One explanation could be a need forcontinuous neural recalibration to opti-mize performance if, as proposed in ourmodel, synapses are always changing.Thus, over time, there might be a slightdrop-off in performance on the basis ofsynaptic turnover alone. For an expert whoperforms at the highest level, even a slightdecrement in performance can be obviousand exploited by competition, therebymaking practice immediately before anevent required to ½ne-tune the network in -to a state of optimal performance.

Whether or not the proposed model iscorrect in its explanation remains to be de -termined. But what we know for sure isthat the continued use of modern imag-ing technology to probe synaptic dynamicswill provide crucial data in the years aheadto constrain and inform our efforts at un -derstanding the neural processes that un -der lie motor learning.

In this review we have focused on the hardscienti½c questions involved in un der -stand ing the seemingly effortless genera-tion of voluntary movements. With re spectto the peripheral motor system (spi nal cordand muscles), we have pointed out themany dif½culties associated with con trol -ling millions of muscle ½bers partitionedacross dozens of muscles, and de scribedhow, through spinal cord modularity, thecns has found a simplifying solution.How ever, no answers have yet been foundto explain how the cortical motor areas ofthe frontal lobe construct the spatiotem-poral patterns of neural activity necessaryto activate the spinal cord, enabling it toexecute a speci½c movement. Certainly, wedo know that high-level movement goalsand attention-related signals are repre-sented in the premotor areas and that thespread of these signals to the primary mo -tor cortex, possibly already primed by af -ferent information about limb posture, will

somehow trigger the retrieval of motormemories and, subsequently, the forma-tion of a signal to the spinal cord. But thedetail of this complicated pro cess, whichcritically involves coordinate and variabletransformations from spatial movementgoals to muscle activations, needs to beelab orated further. Phrased more fanci-fully, we have some idea as to the intricatedesign of the puppet and the puppet strings,but we lack insight into the mind of thepup peteer.

We have also discussed the hard prob-lem of where and how motor memoriesare stored. This, too, is a dif½cult and un -solved problem, in large measure becauseof the highly distributed and interconnec -ted nature of neural circuits. Based on ½rstprinciples, we can be sure that memorystorage in the brain, however it works, willdiffer radically from information storagein a computer. New computational para-digms may be needed to provide a greaterunderstanding, and we have here brieflydescribed one model that attempts to shiftthe paradigm based on our knowledge ofperpetually fluctuating synapses.

On the face of it, investigating the prob -lems of both pattern formation (for thepur pose of control) in the motor cortexand motor memory storage in the aggre-gate motor circuit is going to be a dauntingaffair requiring the combined efforts ofphysiologists, molecular biologists, andcomputational neuroscientists. But as thehistory of science has shown, it is possiblethat nature might have developed somesur prisingly simple and unexpected short -cuts that, if discovered, may go a long waytoward providing the ultimate answers.Af ter all, who would have guessed that asimple receptive ½eld in the visual cortexmight be the key to recognizing the com-plexities of a face?

94

Under -standing

VoluntaryMovements


endnotes* Contributor Biographies: EMILIO BIZZI, a Fellow of the American Academy since 1980 and

President of the Academy from 2006 to 2009, is Institute Professor in the Department ofBrain and Cog nitive Sciences and Investigator at the McGovern Institute for Brain Researchat the Massachusetts Institute of Technology. His recent publications include articles in suchjournals as Frontiers in Computational Neuroscience, Proceedings of the National Academy of Sci-ences, Neuron, and the Journal of Neurophysiology.

ROBERT AJEMIAN is a Research Scientist at the McGovern Institute for Brain Research at theMassachusetts Institute of Technology. His publications include articles in such journals asNeuron, Cerebral Cortex, the Journal of Motor Behavior, and the Journal of Neurophysiology.

Authors’ Note: We would like to acknowledge the following grants, which funded re searchthat contributed to this essay: National Science Foundation Grant IIS-0904594, Program forCollaborative Research in Computational Neuroscience; and Na tional Institutes of HealthGrants (ninds) NS44393 and NS068103.

1 Emilio Bizzi and Vincent C. K. Cheung, “The Neural Origin of Muscle Synergies,” Frontiers inComputational Neuroscience 7 (51) (2013): 1, doi:10.3389/fncom.2013.00051.

2 Emilio Bizzi, Vincent C. K. Cheung, Andrea d’Avella, Philippe Saltiel, and Matthew Tresch,“Combining Modules for Movement,” Brain Research Reviews 57 (2008): 125–133.

3 Ibid.4 Matthew C. Tresch, Vincent C. K. Cheung, and Andrea d’Avella, “Matrix Factorization Algo-

rithms for the Identi½cation of Muscle Synergies: Evaluation on Simulated and ExperimentalData Sets,” Journal of Neurophysiology 95 (4) (2006): 2199–2212, doi:10.1152/jn.00222.2005.

5 Bizzi and Cheung, “The Neural Origin of Muscle Synergies.”6 The cortex, the outer layer of the cerebrum, is composed of folded gray matter and plays an

important role in consciousness.7 A dendrite is a short branched extension of a nerve cell, along which impulses received from

other cells at synapses are transmitted to the cell body.8 Ariel J. Levine, Christopher A. Hinckley, Kathryn L. Hilde, Shawn P. Driscoll, Tiffany H. Poon,

Jessica M. Montgomery, and Samuel L. Pfaff, “Identi½cation of a Cellular Node for MotorControl Pathways,” Nature Neuroscience 17 (4) (2014): 586–593, doi:10.1038/nn.3675.

9 Jean-Alban Rathelot and Peter L. Strick, “Subdivisions of Primary Motor Cortex Based onCor tico-Motoneuronal Cells,” Proceedings of the National Academy of Sciences 106 (3) (2009):918–923, doi:10.1073/pnas.0808362106.

10 E. V. Evarts, “Relation of Pyramidal Tract Activity to Force Exerted During Voluntary Move-ment,” Journal of Neurophysiology 31 (1) (1968): 14–27.

11 Apostolos P. Georgopoulos, John F. Kalaska, Roberto Caminiti, and Joe T. Massey, “On theRelations between the Direction of Two-Dimensional Arm Movements and Cell Discharge inPri mate Motor Cortex,” The Journal of Neuroscience 2 (11) (1982): 1527–1537.

12 Stephen H. Scott, “Inconvenient Truths about Neural Processing in Primary Motor Cortex,”The Journal of Physiology 586 (5) (2008): 1217–1224, doi:10.1113/jphysiol.2007.146068

13 For the optimal feedback control model, see Emanuel Todorov and Michael I. Jordan, “OptimalFeedback Control as a Theory of Motor Coordination,” Nature Neuroscience 5 (11) (2002):1226– 1235, doi:10.1038/nn963; and for the recurrent neural network model, see Mark M.Church land, John P. Cunningham, Matthew T. Kaufman, Justin D. Foster, Paul Nuyujukian,Stephen I. Ryu, and Krishna V. Shenoy, “Neural Population Dynamics during Reaching,” Nature487 (7405) (2012): 51–56.

95


144 (1) Winter 2015

14 Chiang-Shan Ray Li, Camillo Padoa Schioppa, and Emilio Bizzi, “Neuronal Correlates of MotorPerformance and Motor Learning in the Primary Motor Cortex of Monkeys Adapting to anExternal Force Field,” Neuron 30 (2) (2001): 593–607.

15 Tonghui Xu, Xinzhu Yu, Andrew J. Perlik, Willie F. Tobin, Jonathan A. Zweig, Kelly Tennant,Theresa Jones, and Yi Zuo, “Rapid Formation and Selective Stabilization of Synapses forEnduring Motor Memories,” Nature 462 (2009): 915–919, doi:10.1038/nature08389; and GuangYang, Feng Pan, and Wen-Biao Gan, “Stably Maintained Dendritic Spines are Associated withLifelong Memories,” Nature 462 (2009): 920–924, doi:10.1038/nature08577.

16 Robert Ajemian, Alessandro D’Ausilio, Helene Moorman, and Emilio Bizzi, “A Theory for HowSensorimotor Skills are Learned and Retained in Noisy and Nonstationary Neural Circuits,”Proceedings of the National Academy of Sciences 110 (52) (2013): E5078–E5087, doi:10.1073/pnas.1320116110; and Robert Ajemian, Alessandro D’Ausilio, Helene Moorman, and EmilioBizzi, “Why Professional Athletes Need a Prolonged Period of Warm-Up and Other Peculiaritiesof Human Motor Learning,” Journal of Motor Behavior 42 (6) (2010): 381–388.

a hard scienti½c quest: understanding voluntary movementsweb.mit.edu/ajemian/www/daedalus_2015.pdfa...

Documents