Exp Brin Res (26) 173: 425 437 DOI 1.17/s221-6-391- RESEARCH ARTICLE Eun Jung Hwng Æ Murice A. Smith Rez Shdmehr Dissocible effects of the implicit nd explicit memory systems on lerning control of reching Received: 22 July 25 / Accepted: 25 Jnury 26 / Published online: 28 Februry 26 Ó Springer-Verlg 26 Abstrct Adptive control of reching depends on internl models tht ssocite sttes in which the limb experienced force perturbtion with motor commnds tht cn compenste for it. Limb stte cn be sensed vi both vision nd proprioception. However, dpttion of reching in novel dynmics results in generliztion in the intrinsic coordintes of the limb, suggesting tht the proprioceptive sttes in which the limb ws perturbed dominte representtion of limb stte. To test this hypothesis, we considered tsk where position of the hnd during rech ws correlted with ptterns of force perturbtion. This correltion could be sensed vi vision, proprioception, or both. As predicted, when the correltions could be sensed only vi proprioception, lerning ws significntly better s compred to when the correltions could only be sensed through vision. We found tht lerning with visul correltions resulted in subjects who could verblly describe the ptterns of perturbtions but this wreness ws never observed in subjects who lerned the tsk with only proprioceptive correltions. We mnipulted the reltive vlues of the visul nd proprioceptive prmeters nd found tht the probbility of becoming wre strongly depended on the correltions tht subjects could visully observe. In ll conditions, wre subjects demonstrted smll but significnt dvntge in their bility to dpt their E. J. Hwng Æ M. A. Smith Æ R. Shdmehr (&) Lbortory for Computtionl Motor Control, Deprtment of Biomedicl Engineering, Johns Hopkins School of Medicine, 419 Trylor Building, 72 Rutlnd Ave, Bltimore, MD 2125, USA E-mil: rez@bme.jhu.edu Tel.: +1-41-6143424 Fx: +1-41-614989 Present ddress: E. J. Hwng Division of Biology 216-76, Cliforni Institute of Technology, Psden, CA 91125, USA motor commnds. Proprioceptive correltions produced n internl model tht strongly influenced reching performnce yet did not led to wreness. Visul correltions strongly incresed the probbility of becoming wre, yet hd much smller but still significnt effect on reching performnce. Therefore, prctice resulted in cquisition of both implicit nd explicit internl models. Keywords Reching Æ Arm movements Æ Awreness Æ Adpttion Æ Force fields Æ Vision Æ Proprioception Æ Computtionl models Æ Motor control Æ Motor lerning Introduction The brin cn sense the stte of system vi multiple sensory chnnels. For exmple, both visul nd proprioceptive sensors often simultneously encode rm position (Rossetti et l. 1995). Both visul nd uditory chnnels my encode loction of n object (Cohen nd Andersen 2). It hs been suggested tht when such duplicte informtion is vilble, the brin forms single estimte by weighting ech modlity ccording to its mesurement precision (Ernst nd Bnks 22; vn Beers et l. 1999). However, recent study chllenged this view. Sober nd Sbes (23) considered reching tsk where stte of the limb could be estimted using both vision nd proprioception. They described two importnt psychophysicl results: first, the estimte of limb stte used for kinemtic plnning of rech relied mostly on representtion of limb stte in terms of vision. This is consistent with the ide tht the pln for reching movement forms in the posterior prietl cortex in terms of vectors tht represent hnd nd trget positions with respect to fixtion (Buneo et l. 22). The difference between these two vectors is vector tht points from the hnd to the trget. In the premotor cortex, the representtion of this vector remins lrgely independent of the proprioceptive stte of the limb
426 (Kkei et l. 21). Second, the estimte used for trnsforming the plnned movement into motor commnds relied mostly on the limb s proprioceptive stte. This is consistent with the result tht for given movement direction, proprioceptive stte of the rm strongly ffects dischrge of motor corticl cells (Scott nd Klsk 1997). Therefore, it is plusible tht visul nd proprioceptive informtion regrding limb stte differentilly ffect plnning versus execution of rech. When there re novel forces on the hnd tht depend on stte of the limb, for exmple while reching in force field (Shdmehr nd Muss-Ivldi 1994), the brin lerns to lter the mps tht describe the execution of the rech nd not the mps tht describe rech plns. For exmple, when people prctice reching in force field, it ppers tht the brin ssocites the limb stte in which forces were experienced to the motor commnd tht re necessry to negte the forces (Conditt et l. 1997). The coordinte system in which limb stte is represented in this computtion is joint-like rther thn Crtesin, resulting in generliztion ptterns tht re in the intrinsic coordintes of the rm (Shdmehr nd Moussvi 2; Mlfit et l. 22). This pttern of generliztion is consistent with the hypothesis tht the brin relies primrily on proprioception to represent limb stte for the purpose of trnsforming movement pln into motor commnds. Two recent results, however, re inconsistent with this view. An internl model of dynmics tht represents stte of the rm in terms of proprioception predicts tht when the rm experiences pttern of force during reching, there should be no trnsfer of lerning to the contrlterl rm. However, there is smll but significnt generliztion from one rm to the other nd the coordinte system ppers to be Crtesin (Criscimgn-Hemminger et l. 23; Mlfit nd Ostry 24). Moreover, recent study showed tht when n observer views video of n ctor reching in force field, he gins informtion tht hs smll but significnt effect on his performnce (Mttr nd Gribble 25). To reconcile these results, we considered the possibility tht sensory informtion bout stte of the limb might contribute to performnce in two fundmentlly different wys. First, visul nd proprioceptive informtion might combine to form n implicit internl model of dynmics. This implicit model would likely depend on n estimte of limb stte s sensed through proprioception. Second, visully observed ptterns might help form n explicit internl model tht predicts the direction or other chrcteristics of the perturbtion. This might result in wreness. The two components of knowledge would combine to guide performnce. Thus, in this study we ttempted to elucidte three points: (1) the reltive contribution of vision nd proprioception to implicit force field lerning, (2) the reltive contribution of vision nd proprioception to wreness of the force field pttern, nd (3) the role of wreness in force field lerning. The results suggest revision of existing computtionl models of dpttion for reching. Methods Fifty-four helthy individuls (18 women nd 36 men), ll nïve to the experiment, prticipted in this study. Averge ge ws 25 yers (rnge: 19 35 yers). The study protocol ws pproved by JHU School of Medicine IRB nd ll subjects signed consent form. Subjects st on chir in front of 2D robotic mnipulndum nd held its hndle. Their upper rm rested on n rm support ttched to the chir nd they reched in the horizontl plne t the shoulder height. We spred sheet of blck hevy cloth horizontlly bove the movement plne. This sheet occluded view of the entire body below the neck (including the moving rm). A verticl monitor ws plced bout 75 cm in front of subjects nd displyed cursor (2 2 mm 2 ) representing hnd position nd squres (6 6 mm 2 ) representing strt nd trget positions of reching. The reltionship between this visul disply nd hnd position vried in the five experimentl groups (s shown in Fig. 1 nd detiled below). Tsk A cursor on the monitor represented hnd position. It lwys moved with velocity tht ws identicl to hnd velocity. On ech tril, one of three force fields ~F tht depend on the hnd velocity ~_x; where ~F ¼ B ~_x; cted on the hnd: clockwise (CW) curl field (B=[ 13; 13 ] NÆs/m), counter-clockwise (CCW) curl field (B=[ 13; 13 ] NÆ s/m), or null field. These fields pushed the hnd perpendiculr to the direction of motion. As the movements were lwys to trget t 18 (i.e., downwrd), CW field pushed the hnd to the left while CCW field pushed the hnd to the right. We generted 42 long pseudo-rndom sequence of fields nd used the sme sequence in ll subjects. This sequence controlled the strt position of ech rech tril. For exmple, in the mtched group, when the tril indicted CW field, the robot positioned the hnd t left of center to strt the rech tril nd cursor ws displyed t left of center on the screen. When the tril indicted CCW field, the robot positioned the hnd t the right of center nd cursor ws displyed t right of center on the screen. When the tril indicted null field, the robot positioned the hnd t the middle nd the cursor ws lso displyed in the middle of the monitor. In contrst, in the proprioceptive cue group, the robot positioned the hnd t left, center, or right ccording to the field in tht tril but the cursor strting position ws lwys t center. In the visul cue group, the robot lwys positioned the hnd t center but the cursor ws displyed t left, center, or right depending on the field
427 Mtched Proprioception-only Vision-only Flipped strt v R =7 cm v L =7 cm trget strt p R =7 cm p L =-7 cm trget b c Fig. 1 Experiment design nd post experiment interview. Top row displys the cursor trjectory. The bottom row displys the corresponding hnd trjectory. The curved pths represent perturbed movements by the force field, either clockwise or counterclockwise. Mtched group: the cursor position ws ligned with subject s hnd position. Proprioceptive cue group: hnd position vried but not cursor position. Visul cue group: the cursor position vries but not hnd position. Flipped group: the cursor ws displyed on the opposite side of the hnd position. b Assessment of force field wreness. After the completion of the experiment, ech subject ws sked these questions in order to determine whether they were consciously wre of the force field pttern. c Exmples from two subjects. Some subjects were ble to ccurtely describe the force field pttern with digrms or in writing in tht tril. Note, however, tht cursor velocity lwys mtched hnd velocity nd therefore in ll groups deflection of the hnd to the right or left during movement ws ccurtely displced on the screen. The tsk ws to rech to displyed trget (displcement of 1 cm) within 5±5 ms. Trgets were lwys displyed on verticl line, stright down from strt positions, which represented reching directly towrd the subject. Onset of movement ws determined using n bsolute velocity threshold of.3 m/s. Feedbck on performnce ws provided immeditely fter trget cquisition. If the trget ws cquired within 1 ms window round the required movement time (45 55 ms), the trget exploded nd the computer mde sound. If the trget loction ws cquired too slowly or too quickly, the trget turned blue or red, respectively. After this rewrd ws presented the robot moved the subject s hnd to new strt position. During this trnsition period, the cursor feedbck indicting hnd position ws blnked until the hnd ws within 2 cm of the strt position for the next tril. After the hnd becme sttionry within 5 mm of the new strt position for 3 ms, the video monitor displyed the next trget. Experiment 1 We compred performnce in three conditions: in one condition, visul informtion tht indicted the
428 position of the reching movement correlted with the ptterns of force perturbtion during the rech. In nother condition, proprioceptive informtion tht indicted the position of the rech correlted with the pttern of forces. In finl condition, both proprioceptive nd visul informtion correlted with the forces. In the mtched group (n=8) t the strt of ech rech the robot positioned the hnd t one of three loctions left, center, or right (Fig. 1). These loctions were evenly spced (7 cm prt from ech other) on line bout 48 cm in front of the subject nd centered bout the subject s midline. If the field for tril ws CW, then the robot positioned tht hnd t the left strt loction; right if the field ws CCW; center if the field ws null. The cursor ws similrly positioned t left ( 7 cm), center ( cm), nd right (+7 cm) to mtch strting hnd position. In the proprioceptive cue group (n=7) t the strt of ech rech the robot positioned the hnd t left, center, or right depending on the field condition for tht tril (sme hnd position-field correspondence s in the mtched group). However, the cursor displyed the strting hnd position lwys t the center. When the movement begn, the velocity vector for the cursor mtched hnd velocity. Tht is, if the hnd ws pushed to the right, the cursor moved to the right by exctly the mount tht the hnd ws pushed. This mtching of cursor velocity nd hnd velocity ws mintined in ll conditions for ll groups. In the visul cue groups (n=7), t the strt of ech rech the robot lwys positioned the hnd t the middle of the workspce. However, the cursor ws displyed t left, center, or right depending on the field condition for tht tril (sme cursor position-field correspondence s in the mtched group). The cursor distnce between left nd right movements ws 7 cm. Experiment 2 Results of Experiment 1 suggested to us tht when there ws correltion between proprioceptively sensed movement positions nd perturbtions, subjects lerned the tsk well. However, when this correltion ws sensed through vision, there ws good chnce of becoming wre of the ptterns of perturbtion. Did wreness hve significnt impct on performnce? To nswer this question, we recruited lrge number of new subjects into flipped group (n=16) where t the strt of ech rech the robot positioned the hnd t left, center, or right ccording to the field in tht tril. However, the cursor tht indicted strting hnd position ws displyed opposite to the loction of the hnd. For exmple, when the hnd ws resting t the right strting position, the cursor ws displyed to left of the monitor midline. As before, once the movement begn the cursor velocity mtched hnd velocity. Experiment 3 The results of Experiment 2 demonstrted tht wre subjects performed significntly better thn unwre subjects. To determine the extent of this dvntge, we considered condition where we expected tht nerly ll subjects should become wre. We gin recruited new group of subjects (n=12) nd trined them in prdigm identicl to the visul cue group described bove, except tht the cursor indicting hnd position ws t 14 cm to the left or right of the center of the screen. As before, in ll trils the robot plced the hnd t the center strt loction. Experiment 4 Bsed on the results of Experiments 1 3, we constructed model. To directly test the predictions of our model we considered sitution where subjects were instructed explicitly regrding the reltionship between strt position of the rech nd the ptterns of force. In the wre proprioceptive cue group (n=4), we informed subjects of this reltionship in the beginning of the experiment nd conducted the experiment with the sme condition s the proprioceptive cue group. Generl experiment procedures To fmilirize the volunteers with the tsk, subjects begn with three sets of 84 movements in the null field. Following this, subjects did five force field sets of 84 movements in which fields were pplied s described bove. During trining in the field, occsionl pseudorndomly designted trils were performed without forces (i.e., ctch trils). Approximtely one in seven trils ws ctch tril. Note tht the wre proprioceptive cue group subjects performed 8 more movements in the null field during which they could prctice to distinguish the three different strt positions before the min experiment strted. Also, only in the wre proprioceptive cue group, during the null field sets only, words indicting the strt position, left, center, or right were displyed t the beginning of ech tril to help subjects differentite the three different strt positions. After the experiment, subjects completed written questionnire (Fig. 1b). The form contined questions bout self-evlution of performnce, notice of ny force field ptterns, use of ny strtegy to succeed in the tsk, nd level of ttention during the experiment. The purpose of the questionnire ws to help us ctegorize subjects into wre or unwre groups. All subjects who nswered yes to the first question, did you notice ny reltionship between the force direction nd movement strt position?, nd correctly described the force pttern (s in the box in Fig. 2) were ssigned to
Fig. 2 Implicit lerning depends primrily on proprioception but wreness improves lerning. Hnd pths of typicl subject in the vision, proprioception nd mtched groups. Top pnel shows the first field (blck) or ctch (gry) trils nd bottom pnel shows the lst field or ctch trils for ech loction. b Results of Experiment 1. Lerning index of the vision, proprioception, nd mtched groups, with sub-groups indicting wreness clssifiction bsed on postexperiment written questionnire. The index is n verge of performnce over ll 5 sets of trining. See Appendix for dt for ech set. The number inside of ech br indictes the number of subjects. Error brs re stndrd errors of men. c Results of Experiment 2, flipped condition. d Results of Experiment 3, vision:14 cm condition b Lerning index.4.35.3.25.2.15.1.5 field tril Vision-only Proprioception-only Mtched First field (ctch) trils First field (ctch) trils First field (ctch) trils ctch tril Lst field (ctch) trils Lst field (ctch) trils Lst field (ctch) trils Left Center Right Left Center Right Left Center Right vision:7cm 2 u 5 proprioception 6 mtched 1 7 u u c.4.35.3.25.2.15.1.5 flipped 9 u 7 d.4.35.3.25.2.15.1.5 vision:14cm 1 u 11 429 the wre group. Only one subject mrked no to the first question nd then nswered yes to the second question, did you relize tht when the movement strts on the right, force direction ws rightwrd nd when the movement strts on the left, force direction ws leftwrd?. This subject ws lso ctegorized wre. Some subjects becme consciously wre of the pttern of forces. These subjects were ble to describe or confirm the correct reltionship between upcoming force direction nd the strt position of movement wheres unwre subjects either incorrectly described the reltionship or were unble to confirm the correct reltionship (Fig. 1c). When the subjects were sked to describe the reltionship between the force direction nd the strt position, most subjects drew digrm. The typicl digrm included boxes t the three strt loctions nd subsequent hnd pths. The hnd pths were curved towrd the force direction often with rrows indicting the force direction (Fig. 1c). Some subjects reported verblly, e.g., when presenting trgets on the right, the robot pushed right, when presenting trgets on the left, the robot pushed left nd there ws no resistnce in the middle. Performnce mesures As mesure of error, we used the displcement perpendiculr to trget direction t 25 ms into the movement (perpendiculr error, p.e.). Using p.e. t 25 ms, we computed lerning index (Smith nd Shdmehr 25): when lerning occurs, the trjectories of field trils become stright while the trjectories of ctch trils become skewed. Therefore, the error in field trils decreses nd the error in ctch trils increses. We quntified the lerning index s: Lerning Index ¼ p:e: ctch tril =ðp:e: ctch tril p:e: field tril Þ For ech set, men of the perpendiculr errors for the ctch trils nd the field trils t ech loction ws used to compute the lerning index for the left nd right
43 movements seprtely nd the men of these two indices were tken s lerning index for given set. For ech subject, we verged these indices cross the five dpttion sets to use s single subject s overll lerning index. Sttisticl tests To test the sttisticl significnce in performnce difference mong multiple groups, we performed n onewy or two-wy ANOVA followed by Tukey s Lest Significnt Difference(TLSD) multicomprisons using Mtlb sttisticl toolbox. Becuse the number of subjects in ech group is not equl, we performed n unblnced ANOVA using function, novn, in the Mtlb toolbox. Moreover, the function novn llowed us to test only the min effects using liner model nd enbled us to do n ANOVA even in the cse tht one of the cells in the design mtrix ws empty. In the post-hoc comprisons, we considered P<.5 to be sttisticlly significnt. Results Experiment 1: proprioceptively sensed stte of the limb dominted dptive control In the proprioceptive-only condition, visul cues indicted tht the reches were performed in the sme loction while proprioception indicted tht the rech took plce in three different loctions (seprted by 7 cm). The proprioceptively sensed positions correlted with force perturbtions. In the visul cue condition, proprioception indicted tht the reches were performed in the sme loction while vision indicted tht they took plce in three different loctions (seprted by 7 cm). The visully sensed positions correlted with force perturbtions. In the mtched-group, both visul nd proprioceptive cues indicted tht the reches were performed in three different loctions (seprted by 7 cm). If the brin dptively controls reching by forming n ssocition between proprioceptively sensed positions of the limb where perturbtions were experienced nd the motor commnds tht cn overcome them, then performnce of mtched nd proprioceptive cue groups should be similr while performnce of proprioceptive cue group should be significntly better thn the visul cue group. While we quntified the effect of vision nd proprioception on performnce, we lso considered the possibility tht performnce depended not just on n implicit internl model, but lso on n explicit component. We tried to guge this explicit component vi wreness of the reltionship between the end-effector position nd ptterns of force. We provided written interview to ech subject immeditely fter completion of the tsk (detils described in Methods nd Fig. 1b). Subjects were clssified s wre if they were ble to describe or confirm the correct reltionship between strting positions nd force directions (Fig. 1c). Figure 2 displys hnd pths of typicl subject from the mtched, visul cue, nd proprioceptive cue groups. In the first set of field trils, hnd pths were strongly disturbed by forces in the left nd right loctions in ll three groups. After trining, in the mtched nd proprioceptive cue group the field trils were strighter nd the ctch trils were curved in the opposite direction of the fields, indicting tht subjects formed internl models for the externl perturbtions. However, in the vision only-group trining resulted in smller chnges in both field nd ctch trils thn the other two groups. We quntified performnce for ech subject using lerning index (see Methods). Figure 2b shows verged performnce cross subjects in the three experimentl conditions. [Appendix includes the detiled time course of this index for ech subject group.] We found tht group membership significntly ffected performnce (one-wy ANOVA, F(2,18)=13.2, P<1 3 ). The TLSD post-hoc multicomprison reveled tht the performnce difference ws significnt between the visul cue nd proprioceptive cue groups, visul cue nd mtched groups, nd proprioceptive cue nd mtched groups (See Methods). Performnce in the proprioceptive cue group ws significntly better thn the visul cue group. Therefore, if the subjects could not see the reltionship between movement positions nd perturbtions, but could proprioceptively sense this correltion, they lerned the tsk significntly better thn if the correltion could be sensed vi vision only. Performnce of the mtched group ws significntly better thn the proprioceptive cue group. Therefore, when the correltion between movement positions nd perturbtions could be sensed vi both vision nd proprioception, there were performnce gins with respect to when only one of these modlities ws vilble. The difference between the mtched nd proprioceptive cue group is n estimte of the influence of visul informtion. These results reject our hypothesis tht dpttion ws due to n ssocition of proprioceptively sensed stte of the limb with motor errors. Rther, the results suggest tht both cues were ffecting performnce but tht proprioception ppered to hve dominnt role. We next exmined the performnce of the wre nd unwre subjects in ech group nd mde two observtions. First, mjority of subjects in the vision nd mtched groups gined wreness while no subject becme wre in the proprioception group. This differences in the probbility of wreness in the three groups re significnt (v 2 test, v 2 =12.71, df=2, P<.1). This hinted tht wreness might hve been driven by the visul cues tht signled correltion between movement positions nd perturbtions. Second, we noted tht wre subjects tended to outperform their unwre counterprts. Unfortuntely, this tendency could not be
431 confirmed sttisticlly in within-group comprison becuse the visul cue nd mtched groups contined very few unwre subjects wheres the subjects in the proprioceptive cue group were ll unwre. Becuse in humns, wreness is fundmentl property of the declrtive memory system, the results suggested to us tht in the wre subjects we were observing the combined influence of the explicit nd implicit memory systems. However, we needed to test whether wreness hd significnt effect on performnce. Experiments 2 nd 3: wreness significntly improved performnce In Experiment 2, we recruited new subjects (n=16) nd tested them in new condition. In the flipped group the robot positioned the hnd t left, center, or right. However, when the hnd ws t left, the cursor ws displyed t right nd when the hnd ws t right, the cursor ws displyed t left. Therefore, both visul nd proprioceptive sense of limb position correlted with the ptterns of perturbtion. Figure 2c summrizes the results of the experiment. We found tht seven out of sixteen subjects becme wre of the force pttern nd wre subjects performed significntly better thn unwre subjects (one sided t test, t=2.52, df=14, P=.1). Interestingly, mong the unwre sub-group performnce ws very similr to the performnce of proprioceptive cue group where subjects were lso unwre. We performed two-wy ANOVA for ll the experimentl groups so fr (visul cue, proprioceptive cue, mtched nd flipped groups) with the first fctor being experimentl condition nd the second fctor being wreness. We found tht both experimentl condition nd wreness were significnt fctors (F(3,33)=9.54, P<1 3 nd F(1,33)=8.48, P<.1). The TLSD post-hoc multicomprisons showed tht the performnce ws sttisticlly different between the visul cue nd proprioceptive cue groups, visul cue nd mtched groups, nd visul cue nd flipped groups. Performnce of the unwre flipped group ws not significntly different from the unwre proprioceptive cue group. Therefore, when proprioceptive feedbck ws identicl, performnce of unwre subjects ws similr (i.e., proprioceptive cue, mtched-unwre, nd flipped-unwre subgroups). Notice tht the performnce difference between wre nd unwre subgroups in visul cue, mtched, nd flipped groups is nerly constnt. To further test this, we performed two-wy ANOVA for the visul cue, mtched nd flipped groups using n interction model nd found no significnt interction effect. Thus, wreness ppered to improve performnce by roughly constnt mount regrdless of the experimentl condition. To quntify the effect of wreness further, we recruited new subjects (n=12) nd tested them in visul cue condition. On every tril the robot plced the hnd t the center strt position but the cursor ws displyed t left, center, or right (distnce from center=14 cm). As before, the pttern of force perturbtions depended on where the cursor ws plced on the screen. We found tht 11 out of 12 subjects becme wre. Performnces of the two sub-groups re plotted in Fig. 2d. In every group where subjects becme wre, performnce of the wre sub-group ppered better thn their unwre counterprts. An dditive model of the implicit nd explicit components of lerning To help explin these results, we mde model bsed on two ssumptions. (1) Suppose tht the unwre (implicit) lerning nd wre (explicit) lerning ech depend on the brin s bility to estimte limb position in ech tril. The estimte of limb position depends on the vlues observed by vision nd proprioception, nd lso the confidence tht the brin hs in those vlues. (2) Suppose tht the performnce tht we mesure is sum of contributions mde by the implicit (unwre) nd explicit (wre) memory systems. In tril i, the distnce of the visul cue from the midline is represented by v (i), nd distnce of the proprioceptive cue from the midline by vrible p (i). If the confidence in the visul cue is noted with c v nd the confidence in the proprioceptive cue is noted with c p, then the perceived stte of the limb q (i) is (Ernst nd Bnks 22): q ðiþ ¼ c vv ðiþ þ c p p ðiþ ð1þ c v þ c p Erlier we hd reported tht performnce (i.e., lerning index) s depended linerly on the distnce between the left nd right movements (Hwng et l. 23). Thus, we hve: s ¼ k q ðrþ q ðlþ ð2þ For exmple, Eqs. 1 nd 2 explin tht performnce in the mtched group ws better thn the proprioceptiveonly group becuse in the mtched group the visul feedbck confirmed the proprioceptive feedbck, thereby producing lrger perceived distnce between the movements, q (r) q (l). Eqs. 1 nd 2 include only the effect of seprtion distnce between left nd right movements nd not the effect of wreness. Suppose tht if subject is wre of the force ptterns, then the resulting explicit knowledge would contribute n mount e to the lerning index. Performnce of unwre nd wre groups now becomes: s unwre ¼ kðq ðrþ q ðlþ Þ¼w v ðv ðrþ v ðlþ Þþw p ðp ðrþ p ðlþ Þ ð3þ s wre ¼ s unwre þ e ð4þ
432 Here, w v ¼ kc v =ðc v þ c p Þ; w p ¼ kc p =ðc v þ c p Þ; nd e is constnt indicting the mount of performnce improvement due to wreness (Fig. 3). Eq. 4 indictes tht in every group, wre subjects should perform better thn unwre subjects. It lso indictes tht performnce improvement due to wreness should be similr in ll groups. We fitted the three unknown prmeter of Eqs. 3 nd 4 to the men of the group dt (including seprte sub-groups for wre nd unwre) nd found n r 2 =.93 (Fig. 3b). Therefore, the model ccounted for the group dt exceedingly well. How well did the model ccount for individul subjects? Using the sme model, we computed n r-squre for dt from individul subjects in Experiments 1 3 (Fig. 3b) nd found highly significnt fit (P<.1, r 2 =.27). The fit estimted tht w v =.27±.42 (men±95% confidence), w p =.148±.57, nd e=.14±.1. This resulted in c v =.16, nd c p =.84. Therefore, performnce of unwre subjects ws significntly ffected by proprioceptive informtion, while performnce in ll subjects ws significntly ffected by wreness. Experiment 4: providing explicit knowledge vi instructions improved performnce Although our dt suggests tht the wre subjects outperform their unwre counterprts by roughly constnt mount s our model s fit to the dt shows, the smll number of subjects in the unwre subgroup in ech experimentl condition except the flipped condition mkes it hrd to prove this point. To test how well this model predicted performnce on new dt, we recruited yet nother group of subjects. Our model predicted tht if subjects in the proprioceptive cue group becme wre of the field pttern, they would perform better thn the unwre proprioceptive cue subjects by similr mount s the other conditions. The blck dot in Fig. 3b indictes the model s prediction. To test this prediction, we needed wre proprioceptive cue subjects. However, we hd erlier found tht it ws unlikely tht subjects in the proprioceptive cue condition would become wre. Thus, we recruited four new subjects nd before beginning of experiment used digrm to inform these subjects of the reltionship between the strt position of Fig. 3 Model fit of lerning performnce nd probbility of wreness. The model of unwre nd wre lerning described in Eqs. (3) nd (4). b The dt (squres, representing men±sem) were fit to Eqs. (3) nd (4). The term e indictes improvement of lerning due to wreness. The model fit (Eqs. 3, 4) to the group dt ws r 2 =.93. The model fit to the individul subjects ws r 2 =.27, P<.1 when k=.15, e=.1 nd c v =.16, c p =.84. Note tht the wre proprioceptive cue group (blck dot) ws not used to fit the model, but rther ws the model s prediction tht we tested in Experiment 4. The gry circle round the blck dot is the ctul dt for Experiment 4. c The model fit to probbility of wreness, where q is perceived stte s in Eq. (5). Prmeter vlues: c v =.8, nd c p =.2 b Lerning index.4.3.2.1 p (r) -p (l) q' (r) -q' (l) cp cv+cp q (r) -q (l) + + e=.1 v (r) -v (l) cv cv+cp wre unwre 2 4 6 8 1 12 14 16 distnce between left nd right strt positions = q(r)- q(l) :vision 7cm :vision 14cm :flipped :proprioception :mtched θ k c probbility of force-field wreness 1.9.8.7.6.5.4.3.2.1 probbility of wreness s wre flip mtch vision 7 vision 14 proprioception 4 8 12 16 2 24 28 distnce between left nd right strt positions = q' (r) - q' (l)
433 ech rech nd the perturbtion ptterns. The gry circle in Fig. 3b shows the performnce of these subjects. Their performnce ws essentilly identicl to the model s predicted performnce, confirming nerly constnt improvement due to wreness. Providing explicit knowledge vi instructions significntly improved performnce (one-sided t test, t=2.25, df=9, P=.25). To summrize the effect of the experimentl condition nd wreness, we performed two-wy ANOVA cross ll the experimentl groups including the wre proprioceptive cue group nd found tht both the experimentl condition nd wreness were significnt fctors (F(4, 48)=3.7, P<.1 nd F(1,48)=5.94, P<.5). The post-hoc multicomprisons showed tht the performnce difference ws significnt between the visul cue nd proprioceptive cue groups, visul cue nd mtched groups, visul cue nd flipped groups, nd mtched nd double visul cue groups. Probbility of becoming wre depended minly on visul informtion Finlly, we considered how the probbility of becoming wre Pr () depended on visul nd proprioceptive informtion. We represented wreness s binry vrible nd used Pr () to represent the probbility of becoming wre. We imgined tht this probbility my depend on performnce, i.e., the better the performnce, the higher the probbility of becoming wre. In such scenrio, wreness would be triggered when subject s performnce reched certin threshold. Our dt rejected this possibility: In the proprioceptive cue group, where performnce ws high, Pr ()=, wheres in the vision:7 cm group, where performnce ws poor, Pr ()=.71. We next considered the ide tht the probbility of becoming wre depended on the presence of visul nd proprioceptive informtion nd not performnce, Pr ()=g(v, p). Thus, in group of subjects the expected vlue of the lerning index should depend on both the unwre lerning nd the probbility of wreness: s totl ¼ s unwre þ PrðÞe Figure 3c plots Pr () for ll groups. In the vision:7 cm nd vision:14 cm groups proprioceptive cues were identicl but Pr () incresed from.72 to.92. This suggested tht Pr () depended on the visully perceived distnce of the movements. However, proprioception ws lso fctor becuse while visul distnces of the cues were identicl in the flipped nd mtched groups, Pr () incresed from.44 to.875. Therefore, both visul nd proprioceptive informtion ffected wreness. To estimte the reltive importnce of ech cue, we ssumed tht q ws the perceived stte s in Eq. 1, except with new confidence coefficients c 1 nd c 2. Probbility of wreness depended on q vi logistic function: q ðiþ ¼ c v vðiþ þ c p pðiþ c v þ c p h h ii PrðÞ ¼ 1 þ exp h T 1; q ðrþ q ðlþ 1 ð5þ The fit of this model to the dt in Fig. 3c produced vlues: c v =.8, nd c p =.2 (r 2 =.88, P<.5). Thus, probbility of becoming wre ws four times more dependent on visul informtion thn proprioceptive informtion. Discussion The present study exmined whether there were dissocible explicit nd implicit components to the force field lerning prdigm nd quntified the reltive influence of proprioception nd vision on ech component. We considered tsk where reching movements intercted with force fields. The fields depended on strting position of the reching movements. In the visul cue group, strting position vried in the visul spce but not in the proprioceptive spce. In the proprioceptive cue group, strting position vried only in the proprioceptive spce. In the mtched group, strting position vried in both spces congruently nd in the flipped group, strting position vried in both spces but incongruently. We found tht the mjority of subjects becme wre of the force field dependency on the strt position when the visul cue position vried while no subject becme wre when only the proprioceptive cue position vried. Moreover, in ech group, the wre subjects outperformed their unwre counterprts by smll but significnt mount. Also performnce comprison mong visul cue, proprioceptive cue, mtched nd flipped groups fter removing the wreness dependent lerning (explicit lerning) reveled tht implicit lerning system relied primrily on proprioceptive cue. Therefore, we report the new observtion tht in lerning dynmics of reching, the brin relies on both implicit nd explicit lerning systems. Performnce ws dominted by the implicit lerning system nd tht system relied primrily on proprioception to form internl models of dynmics. However, performnce ws significntly ffected by n explicit lerning system tht becme wre of the force ptterns nd this wreness depended primrily on visul informtion in the tsk. A role for wreness in control of reching In mny motor skills tht hve been studied in the lbortory, e.g., prism dpttion, visul rottion, nd seril rection time (SRT) sequence lerning, there re distinct components of performnce tht my be due to wreness. For exmple, in the SRT tsk wreness of the underlying sequence cn develop s consequence of prolonged trining (Stdler 1994) or simply through
434 cue tht signls the introduction of the sequence (Willinghm et l. 22). Ptients with unilterl prietl cortex dmge show no bility to lern the SRT tsk (Boyd nd Winstein 21). However, their performnce improves significntly if they first lern the sequence explicitly. These results not only support the ntomicl distinction between regions tht support explicit nd implicit lerning, but suggest tht explicit knowledge cn help improve motor performnce. A well studied motor lerning tsk is the control of reching in force fields. Subjects with severe impirment in their declrtive memory system, e.g., mnesic ptients, pper to hve norml lerning nd retention in this tsk (Shdmehr et l. 1998). However, this does not imply tht explicit knowledge plys no role in the tsk. In recent study, Osu et l. (Osu et l. 24) demonstrted tht subjects could lern to ssocite force perturbtions with visul cues when they were provided with explicit verbl informtion. However, when such instructions were not provided, extensive trining on similr tsk did not produce ny improvement in performnce (Ro nd Shdmehr 21). Therefore, it is possible tht even in the force field tsk; explicit knowledge might ply role in performnce. Here we found tht wre subjects in every group tended to perform better thn their unwre counterprts. The effect of wreness ppered to be constnt increse in performnce. Note tht the smll number of subjects in the unwre subgroup in ech experimentl condition except the flipped condition mkes it hrd to prove tht there ws constnt mount of performnce improvement. However, the results of the experiment 4 tht wre subjects in the proprioceptive cue condition outperformed their unwre counterprts by the similr mount s the other conditions support the ide tht wreness results in constnt increse in performnce. Then, how did wre subjects perform better? One might expect tht wre subjects developed some strtegies to succeed in the tsk, e.g., iming differently depending on the expected direction of perturbtion. We closely nlyzed kinemtic fetures like hnd pths, speed profile, nd perpendiculr velocity but could not detect n obvious mrker tht distinguished n wre subject (Appendix). Indeed, in our post-experiment survey ll of our wre subjects denied tht they imed differently depending on the expected force field direction. Insted, most wre subjects reported tht they were mentlly getting prepred to resist ginst the expected direction of perturbtion. Effect of vision-proprioception mismtch Our experiment imposed discrepncy between the strt positions of the rech s viewed visully versus felt proprioceptively. Importntly, once the movement begn, the displcements in the two coordinte systems remined consistently ligned. While it is known tht discordnt hnd-cursor motion impirs the subject s visul trcking, discordnce is generlly due to velocity mismtch rther thn bsolute strt position. In our experiment, there ws no velocity mismtch except tht the horizontl movements were projected onto the verticl plne which would ffect ll the experimentl groups. The visul cue nd proprioceptive cue groups might hve performed worse thn the mtched group simply due to the visul-proprioceptive discordnce. According to this generl impirment scenrio, both the flipped nd the 14 cm visul cue groups would perform worse thn the 7 cm visul cue group becuse the former two conditions induced lrger discrepncy, impiring the trcking bility even more thn the 7 cm visul cue group. On the contrry, our dt showed tht subjects in the flipped nd 14 cm visul cue group performed better, suggesting tht mismtch-induced generl impirment is unlikely to hve lrge effect. Reltive contributions of vision nd proprioception Until very recently, mny studies showed tht vision domintes when vision nd proprioception conflict in perceptul tsks (Ernst nd Bnks 22). However, Sober nd Sbes (23) found tht the reltive contributions of vision nd proprioception re not fixed but vrible depending on the computtion being performed. Furthermore, even for the sme computtion, the brin weights differently depending on the context such s the content of the visul informtion or the sensory modlity for the trget loction, suggesting tht the brin tries to minimize errors rising from trnsformtion between coordintes (Sober nd Sbes 25). From this point of view, one reson why we found the visul informtion ws severely discounted in implicit lerning system might be relted to our experimentl setup which used verticl monitor insted of virtul relity or horizontl projector. In our experimentl setup, the visul informtion presented on the verticl monitor needs to be trnsformed to the horizontl plne, nd so it is possible tht the visul cue ws further discounted to reduce errors from this dditionl trnsformtion. Model predictions nd limittions Awreness tht we observed in our tsk likely relies on neurl structures tht re distinct from those tht store implicit knowledge. In this scenrio, the result of trining is formtion of potentilly two different internl models, independently contributing to performnce. This scenrio is intriguing becuse it potentilly cn explin some puzzling recent results. Erlier work hd found tht trining in force field generlizes strongly to the trined rm in intrinsic coordintes but wekly to the opposite rm in extrinsic coordintes (Criscimgn- Hemminger et l. 23; Mlfit nd Ostry 24). If
435 performnce is due to sum of strong implicit internl model tht depends on proprioception nd wek explicit internl model tht depends on vision, then the ptterns of generliztion cn be explined. Furthermore, our model predicts tht if the force field is imposed grdully on the rm so tht there re little or no visul cues to predict the direction of perturbtion, there will be little probbility to become wre, resulting in little or no generliztion to the other rm. This is consistent with recently reported result (Mlfit nd Ostry 24). Finlly, one does not need to ctully move their rm in force field to become wre of the perturbtion ptterns. Viewing of video might suffice. Indeed, recent report found evidence for this ide (Mttr nd Gribble 25). Our model mkes number of novel predictions. We predict tht in mnesic ptients where cquisition of explicit informtion is impired but lerning implicit internl models of reching is unimpired (Shdmehr et l. 1998), there should be little or no trnsfer between rms becuse this trnsfer depends on internl models tht form in explicit memory. Willinghm (Willinghm 21) hs rgued tht the system tht cquires explicit informtion cn guide motor behvior so tht performnce improves, but tht this guidnce requires ttention. Thus, our model predicts tht incresing the ttention lod on reching tsk, e.g., vi distrcter, should: (1) reduce the probbility of becoming wre; (2) reduce performnce with the trined rm only minimlly, but (3) ffect the pttern of generliztion to the untrined rm substntilly. How might wreness influence the neurl representtion of implicit internl models? Consider the bsis function model of force field lerning (Hwng et l. 23) where bses encode the stte of the limb where perturbtion ws sensed. We might imgine tht visully-driven wreness modultes sensitivity of the bses to chnges in rm position. For given sptil distnce between two movements, incresing the sensitivity of the bses to chnges in rm position is equivlent to incresing the seprtion distnce between two movements, thus resulting in improved lerning. Attention or other cues cn indeed modulte tuning of neurons (Reynolds et l. 2; Morn nd Desimone 1985; Musllm et l. 24). For exmple, visul stimuli with low luminnce contrst elicit higher ctivtion of V4 neurons when ttention is directed to the stimulus loction thn when ttention is directed wy from the stimulus loction, demonstrting tht ttention increses the sensitivity to luminnce contrst (Reynolds et l. 2). Thus, if effect of wreness is similr to effect of ttention, sensitivity of the bses to chnges in perceived rm position might be enhnced, resulting in improved lerning. There re limittions on the bility of our study to nswer some of the more importnt questions with regrd to explicit lerning. First, we did not control wreness of explicit knowledge nd rther relied on ech subject s report in the post-experiment survey. It is uncler t which point during the experiment ech subject becme wre. However, when we informed the subjects t the beginning of the trils bout the field-cue pttern (i.e., wre proprioceptive cue group), performnce ws remrkbly close to the model s prediction. This suggests tht in other groups, subjects probbly becme wre erly in trining. Second, we mde binry ctegoriztion for ech subject, wre or unwre, but wreness might not be n ll or none phenomen. Thus, even mong wre subjects, degree of wreness probbly vried. When we fit our model to the group dt, the model explined 93% of the vrince. When we used the sme prmeters to explin the individul dt, we found highly significnt fit (P<.1), but we could only explin 27% of the vrince. This suggests tht the model is n excellent predictor of group performnce, but fr worse in predicting performnce of individul subjects. The inbility to ccount for the vrince within ech sub-group is probbly reflection of our inbility to quntify wreness beyond binomil distribution. In summry, we find tht dptive control of reching in force fields cn produce both implicit nd explicit knowledge, nd both ffect production of motor commnds. The implicit knowledge strongly ffects performnce nd ppers to depend primrily on limb sttes tht re sensed vi proprioception. Visully sensed correltions between limb stte nd perturbtions cn led to wreness, nd the resulting explicit knowledge hs smll but significnt impct on the motor commnds tht control reching. Acknowledgements This work ws supported by grnts from the NIH (NS37422, NS4633). Appendix We exmined the kinemtic fetures of movements in wre nd unwre groups. To eliminte the kinemtic differences due to different performnce level, we compred two groups with similr lerning performnce, vision 14 cm wre nd proprioception unwre groups. Figure 4 displys the verge hnd pths in the lst trining set from these two groups. The velocity perpendiculr to the movement direction from ech group is lso displyed in Fig. 4b. For the left movement, both the field nd ctch trils in these two groups show very similr hnd pths nd velocity profiles. For the right movement, the wre vision 14 cm group shows less perturbtion in both the field nd ctch trils, indicting lrger stiffness of rm in the vision 14 cm group. But the reltive rtio of perturbtion in the field nd ctch trils remins similr in the two groups. We found no specific differences in kinemtics between wre nd unwre groups when the performnce level ws similr. Figure 5 displys lerning index from the mtchedwre, proprioceptive cue, visul cue-unwre(7 cm), flipped-wre, flipped-unwre groups s function of
436 b perpendiculr velocity (m/s) left center right.1.8.6.4.2 -.2 -.4 -.6 -.8 -.1.5 1..5 1..5 1. time (s) Fig. 4 No kinemtic differences between wre nd unwre groups when t similr performnce level.. Averge hnd pths in the lst trining set from the wre vision 14 cm group (gry) nd unwre proprioception group (blck). Solid lines re field trils nd dotted lines re ctch trils. b. Averge profiles of velocity perpendiculr to the movement direction. The formt is sme s in ().6.5.4.3.2.1 -.1 set. An ANOVA test for the block effect in ech group tht consists of more thn two subjects showed sttisticlly significnt effect t P<.5. Becuse lerning index incresed monotoniclly with the set number, we took the men lerning index cross sets s n overll lerning index for ech subject. References 1 2 3 4 5 Set number mtched wre flipped wre mtched unwre proprioception unwre vision wre flipped unwre vision unwre Fig. 5 Lerning index s function of set number. Error brs represent stndrd errors of men vn Beers RJ, Sittig AC, Gon JJ (1999) Integrtion of proprioceptive nd visul position-informtion: n experimentlly supported model. J Neurophysiol 81:1355 1364 Boyd LA, Winstein CJ (21) Implicit motor-sequence lerning in humns following unilterl stroke: the impct of prctice nd explicit knowledge. Neurosci Lett 298:65 69 Buneo CA, Jrvis MR, Btist AP, Andersen RA (22) Direct visuomotor trnsformtions for reching. Nture 416:632 636 Cohen YE, Andersen RA (2) Reches to sounds encoded in n eye-centered reference frme. Neuron 27:647 652 Conditt MA, Gndolfo F, Muss-Ivldi FA (1997) The motor system does not lern the dynmics of the rm by rote memoriztion of pst experience. J Neurophysiol 78:554 56 Criscimgn-Hemminger SE, Donchin O, Gzznig MS, Shdmehr R (23) Lerned dynmics of reching movements generlize from dominnt to nondominnt rm. J Neurophysiol 89:168 176 Ernst MO, Bnks MS (22) Humns integrte visul nd hptic informtion in sttisticlly optiml fshion. Nture 415:429 433 Hwng EJ, Donchin O, Smith MA, Shdmehr R (23) A ginfield encoding of limb position nd velocity in the internl model of rm dynmics. PLoS Biol 1:29 22 Kkei S, Hoffmn DS, Strick PL (21) Direction of ction is represented in the ventrl premotor cortex. Nt Neurosci 4:12 125 Mlfit N, Ostry DJ (24) Is interlimb trnsfer of force-field dpttion cognitive response to the sudden introduction of lod? J Neurosci 24:884 889 Mlfit N, Shiller DM, Ostry DJ (22) Trnsfer of motor lerning cross rm configurtions. J Neurosci 22:9656 966 Mttr AA, Gribble PL (25) Motor lerning by observing. Neuron 46:153 16 Morn J, Desimone R (1985) Selective ttention gtes visul processing in the extrstrite cortex. Science 229:782 784 Musllm S, Corneil BD, Greger B, Scherberger H, Andersen RA (24) Cognitive control signls for neurl prosthetics. Science 35:258 262 Osu R, Hiri S, Yoshiok T, Kwto M (24) Rndom presenttion enbles subjects to dpt to two opposing forces on the hnd. Nt Neurosci 7:111 112 Ro AK, Shdmehr R. Contentul cues fcilitte lerning of multiple models of rm dynmics. Soc Neurosci Abs 27[32.4]. 21 Ref Type: Abstrct Reynolds JH, Psternk T, Desimone R (2) Attention increses sensitivity of V4 neurons. Neuron 26:73 714 Rossetti Y, Desmurget M, Prblnc C (1995) Vectoril coding of movement: vision, proprioception, or both? J Neurophysiol 74:457 463 Scott SH, Klsk JF (1997) Reching movements with similr hnd pths but different rm orienttion: I. Activity of individul cells in motor cortex. J Neurophysiol 77:826 852 Shdmehr R, Moussvi ZMK (2) Sptil generliztion from lerning dynmics of reching movements. J Neurosci 2:787 7815