Articles in PresS. J Neurophysiol (December 5, 2012). doi:10.1152/jn.00509.

2012
1

5 Interaction of temporal and ordinal representations in movement sequences

7 Katja Kornysheva, Anika Sierk, Jörn Diedrichsen

8 Institute of Cognitive Neuroscience

9 University College London

10

11

12 Running Head: Interaction of sequential movement order and timing

13
14 Corresponding author:
15 Katja Kornysheva, PhD
16 Institute of Cognitive Neuroscience
17 University College London
18 Alexandra House
19 17 Queen Square
20 London WC1N 3AR
21 Phone: 0044-20-76795570
22 Fax: 0044-20-78132835
23 E-mail: k.kornysheva@ucl.ac.uk
24

Copyright © 2012 by the American Physiological Society.

 2

25
26 Abstract

27 The production of movement sequences requires an accurate control of muscle

28 activation in time. How does the nervous system encode the precise timing of these

29 movements? One possibility is that the timing of movements (temporal sequence) is an

30 emergent property of the dynamic state of the nervous system and therefore intimately linked

31 to a representation of the sequence of muscle commands (ordinal sequence). Alternatively,

32 timing may be represented independently of the motor effectors and would be transferable to

33 a new ordinal sequence. Some studies have found that a learned temporal sequence cannot be

34 transferred to a new ordinal sequence, thus arguing for an integrated representation. Others

35 have observed temporal transfer across movement sequences and have advocated an

36 independent representation of temporal information. Using a modified serial reaction time

37 task, we tested alternative models of the representation of temporal structure and the

38 interaction between the output of separate ordinal and temporal sequence representations.

39 Temporal transfer depended on whether a novel ordinal sequence was fixed within each test

40 block. Our results confirm the presence of an independent representation of temporal structure

41 and advocate a non-linear multiplicative neural interaction of temporal and ordinal signals in

42 the production of movements.

43

44

45 Keywords: timing, sequence, serial reaction time task, drift-diffusion model

 3

46 Introduction

47 Skilled motor behaviours require precisely timed patterns of muscle activation.

48 Pianists coordinate finger and arm movements to achieve a desired rhythm, loudness and

49 pitch. In our daily life, we coordinate articulatory effectors to produce speech, or different

50 body parts while performing sports activities such as swimming. How does the nervous

51 system store such skilled movement sequences?

52 One possibility is that the next movement in a sequence is issued depending on the

53 current state estimate of the body (Diedrichsen et al., 2010), e.g. a specific configuration of

54 the limbs. Timing would therefore be inseparable from the movements performed, an

55 emergent property of the state of the system, not represented by a dedicated mechanism (Ivry

56 et al., 2002; Spencer et al., 2009; Conditt and Mussa-Ivaldi, 1999). Evidence for this comes

57 from serial reaction time tasks (SRTT), in which participants respond with movements

58 (typically finger movements) to instructive visual cues presented according to a specific

59 sequence. The difference in response times between sequential and random phases

60 (Robertson, 2007) indicates learning of the sequence of digit movements (here referred to in

61 short as “order”). Several studies have shown that the temporal structure (timing) of a trained

62 finger sequence cannot be transferred to a new or random order of movements (Shin and Ivry,

63 2003; O’Reilly et al., 2008). This would argue that the temporal and ordinal structure of

64 sequential movements is indeed stored as an integrated (combined) representation of these

65 parameters.

66 Alternatively, timing may be stored independently of the order of the movements that

67 it was combined with during training. The defining criterion for an independent (as opposed

68 to a combined) temporal representation is that trained timing can produce performance

69 advantages in the context of a new ordinal sequence of movements. Such training-related

70 transfer effects have been observed and in non-cued sequence production tasks (Ullen and
 4

71 Bengtsson, 2003). Here participants were able to learn and transfer trained timing to a new

72 sequence of key presses. The new movement sequence, with which the trained timing was

73 tested, became quickly familiar, since subjects pressed just one key or produced the same

74 sequence sequence repeatedly. These results suggests that the nervous system stores a

75 temporal representation independently of movement order and can - under certain

76 circumstances - transfer this temporal representation to novel sequences of motor commands.

77 How can then the failure to find evidence for independent temporal representation in

78 some SRTT tasks be explained? Here we address this question by examining how

79 independent ordinal and temporal representations interact during movement production. The

80 three possible types of interactions can be illustrated using a drift-diffusion model of

81 sequential action selection (Ratcliff, 1978; Ratcliff and McKoon, 2007) (Fig. 1; for details cf.

82 Methods). The output stage of the model is a selection layer with one unit for each finger (Z).

83 Once a unit reaches a fixed threshold, the corresponding finger response is executed.

84 At the beginning of SRTT learning, the activity of the selection units is driven solely

85 by the stimulus representation S (reactive responses). However, behavioural and

86 neurophysiological findings suggest that with learning of movement sequences an ordinal

87 representation P of the sequence is formed (Shima and Tanji, 1998; Shin and Ivry, 2003;

88 Ullen and Bengtsson, 2003; O’Reilly et al., 2008). Because this ordinal representation can be

89 learned independent of the temporal structure, it is modelled here by tonically pre-activating

90 the next expected unit, as soon as the preceding response is made, e.g. first the index finger,

91 then the thumb etc. When the signal from the stimulus representation is added to the selection

92 layer, the corresponding unit is already closer to threshold, which leads to faster RTs than

93 before learning.

94 Associating a specific movement order to a specific temporal structure has been shown

95 to produce RT advantages compared to a movement order with a changing temporal structure

 5

96 (Shin and Ivry, 2003; O’Reilly et al., 2008). For learning of temporal structure, we consider

97 three possible architectures. One possibility is that a temporal representation T may be stored

98 together with a particular ordinal structure, effectively amounting to another more precisely

99 timed ordinal representation (combined representation, Fig. 1A). In addition to tonically pre-

100 activating the next expected unit it would phasically boost the next unit at an expected point

101 of cue onset leading to faster RT on top of a merely ordinal representation. The timing of

102 movements would consequently be bound to that specific order of movements.

103 Alternatively, a representation of temporal structure may be formed that is

104 independent of the movements with which it is performed. Such a representation would send

105 phasic signals to all the units to boost motor activity, just before the expected time of cue

106 onset. Consequently, the temporal structure would be transferrable to a new order of

107 movements. This independent temporal representation may influence the selection layer in

108 two ways: In the additive model (Fig. 1B), the activation would combine linearly with the

109 activation coming from the ordinal layer. As a consequence, the independent temporal

110 representation would lead to performance advantages even if the ordinal signal were at zero

111 (no ordinal prediction), since the temporal signal is added to all units and therefore facilitates

112 all responses. In contrast, if the temporal signal combines with the ordinal signal

113 multiplicatively (Fig. 1C), then it would lead only to performance advantages once the new

114 ordinal representation has been formed, since the temporal signal boosts only units that are

115 pre-activated by the ordinal representation.

116 We designed a modified SRTT to identify the type of interaction between temporal

117 and ordinal representations during movement production. Subjects were trained on a

118 particular combination of temporal and ordinal sequence of finger presses by reacting to digits

119 presented on the screen. In the test phase, we tested transfer of learning by recombining the

120 trained temporal and ordinal sequence with a new ordinal and temporal sequence, respectively
 6

121 (Fig. 2B), and measuring the RT advantage compared to a baseline condition which has a new

122 temporal and ordinal structure. Importantly, we introduced conditions in which the test phase

123 allows for learning of the new dimensions through repetition of the sequence (“early phase”:

124 trials 1 to 3; “late phase”: trials 8 to 10 of each test block). If ordinal and temporal sequences

125 were completely integrated (combined representation), we should see no transfer, neither in

126 the early phase nor in the later phase when the new order has become familiar (Fig. 1A, right

127 panel). Alternatively, if the temporal sequence representation were independent of the ordinal

128 representation and interacted additively with the latter, temporal transfer should be observed

129 immediately in the early transfer phase (Fig. 1B). Finally, if ordinal and temporal sequence

130 representations interacted multiplicatively, we should find transfer of order, but no transfer of

131 timing in the early phase. Temporal transfer should emerge towards the late phase once the

132 new ordinal structure has become familiar through repetition (Fig. 1C; Experiment 1).

133 Likewise, it should not be manifest in conditions in which the order of finger presses is

134 random on every trial, since learning of the new sequence is not possible (Experiment 2).

135

136 Materials and Methods

137 Subjects

138 16 neurologically healthy volunteers (8 female, mean age: 25, SD: 3.1) participated in

139 Experiment 1 and 15 in Experiment 2 (7 female, mean age: 21.8, SD: 1.5). There was no

140 overlap in subjects between the two experiments. All subjects were right-handed according to

141 the Edinburgh Inventory of Manual Preference (Oldfield, 1971). None of them were

142 professional musicians. All subjects were naïve concerning the hypothesis of this study. The

143 UCL ethics committee approved experimental and consent procedures.

144

145 Stimuli and Task

 7

146 We implemented a modified version of the Serial Reaction Time Task (Nissen and

147 Bullemer, 1987; Penhune and Steele, 2012). Subjects were seated in front of a computer

148 screen with the five fingers of their right hand placed on a keyboard. Each key had a small

149 groove into which a fingertip could be placed. The transducer underneath each key measured

150 the force of individual fingers. An individually adjustable chin rest was provided for comfort.

151 Subjects were presented with a sequence of white digits (1-5) in the middle of a black screen.

152 The digits 1, 2, 3, 4 and 5 instructed to press the thumb, index, middle, ring and little finger of

153 the right hand, respectively. They were instructed to perform the task as fast and accurately as

154 possible. Each trial started with a warning cue (“!”; duration: 400 ms). It was followed by a

155 sequence of five digits that was timed according to a sequence of five possible inter-stimulus-

156 intervals values (ISI; 600, 800, 1000, 1400, 1700 ms). The first ISI commenced when the

157 warning cue disappeared. Digits remained on the screen until the onset of the next digit or for

158 600 ms after the last digit, respectively. In contrast to a classical SRTT, each trial consisted

159 only of one sequence cycle, starting with the warning cue. The trial duration was 6.5 sec and

160 the inter-trial interval was 7.3 sec (0.8 sec gap between trials). Subjects received feedback on

161 their performance throughout the experiment as follows: If the subjects pressed the correct

162 button within the limits of 50 ms before the onset of the current and 50 ms before the onset of

163 the next digit, that digit turned green. If the response was too early the next digit appeared in

164 yellow. If the response was too late the digit turned turquoise. If the finger press was

165 incorrect, the digit turned red. Subjects received a point only when all digits in a sequence

166 turned green, i.e. when they pressed the right finger at the right moment.

167 Movements were instructed by sequences with a particular combination of digit order

168 and timing (ISI sequence). Digit runs or triplets (e.g. 2-3-4) were excluded from the pool of

169 possible sequences. Identical triplets across sequences were prohibited. The sequences for

170 each subject were randomly generated according to these criteria.

 8

171 Both Experiment 1 and 2 consisted of training and subsequent test phase. The training

172 phase consisted of 20 blocks with 10 trials each. During this phase the subjects were trained to

173 produce a sequence of finger presses (here simply referred to as “order”), in combination with

174 a particular sequence of inter-stimulus-intervals (here referred to as “timing”) (cf. Fig. 2).

175 Subjects were not informed of the separation between training and test phase. However, most

176 subjects reported that they noticed changes to the order and timing of the sequences in the

177 experiment. When asked whether they noticed anything particular about the sequences right

178 after the completion of the experiment, 12 of 16 subjects in Experiment 1 reported that a

179 sequence has been repeated at the beginning of the experiment, with 8 subjects specifically

180 pointing out that the finger order was changed and 8 subjects pointing out that the timing

181 changed occasionally. 11 of 15 subjects in Experiment 2 reported that a sequence has been

182 repeated in the training phase, with 5 subjects specifically pointing out that the finger order

183 changed and 5 subjects pointing out that the timing changed occasionally.

184 Experiment 1 served to quantify transfer of the learned temporal sequence, and the

185 learned ordinal sequence. Therefore, the test phase of Experiment 1 consisted of 12 blocks of

186 10 trials, with three blocks in each of the four conditions: In the combined condition subjects

187 produced the trained sequence, using the same order and timing (Fig. 2). In the timing

188 condition the cues appeared with the same temporal intervals, but indicating a different order

189 of finger presses. The new order was different in each of the three timing blocks, but was

190 repeated across the ten trials of each block. In the order condition, the same sequence of

191 finger presses as the training phase was cued, but in combination with a new sequence of

192 inter-stimulus-intervals. The new temporal sequence was different in each of the three order

193 blocks, but did not change across the ten trials of each block. Finally, the baseline condition

194 cued a sequence of finger presses and inter-stimulus intervals that was different from any

195 other trained or tested condition. Again, the novel combination of finger and temporal

196 sequence was different in each of the three baseline blocks, but was repeated across the ten
 9

197 trials of each block. This condition served as an individual baseline to estimate the transfer of

198 a combined, timing and order representation. The condition with which the test phase started

199 was counterbalanced across 16 subjects as follows: 1, 2, 3 ,4 ; 2, 3, 4, 1; 3, 4, 1, 2; 4, 1, 2, 3 (1

200 – combined, 2 – timing, 3 – order, 4 - baseline).

201 Experiment 2 served to contrast two temporal transfer conditions, in which the trained

202 temporal sequence was paired with a new finger sequence that either was or was not repeated

203 across trials in a test block. Training was the same as in Experiment 1. After the training

204 phase subjects were tested on five conditions, with three blocks in each condition.

205 Accordingly, the test phase in Experiment 2 consisted of 15 blocks. The condition with which

206 the test phase started was counterbalanced across subjects. The five conditions were as

207 follows: The combined, timing and baseline conditions were the same as in Experiment 1.

208 Here the respective sequence was repeated across all 10 trials of each block (cf. above).

209 Importantly, we also introduced a condition with fixed timing and random order (timingR).

210 Here the order changed on every single trial, rather than on every block. As a baseline for this

211 condition, we also tested a condition in which both timing and order were random on each

212 trial (baselineR). Each new repeated or random sequence contained an ordinal or temporal

213 sequence that subjects did not encounter before. The condition with which the test phase

214 started was counterbalanced across 15 subjects as follows: 1, 2, 3, 4, 5 ; 2, 3, 4, 5, 1; 3, 4, 5, 1,

215 2; 4, 5, 1, 2, 3; 5, 1, 2, 3, 4 (1 – combined, 2 – timing, 3 – baseline, 4 – timingR, 5 – baselineR).

216 Data analysis

217 Data analysis and simulations was conducted using Matlab (The Mathworks, Inc.)

218 Reaction times (RTs) for each response were defined as time at which the force of a finger

219 exceeded 0.6 N. Only correct responses were considered. Also, responses that occurred more

220 than 100 ms before stimulus onset or more than 600 ms after stimulus onset were considered

221 as errors and excluded from further analysis. For each trial, we averaged the RT for all correct
 10

222 responses. We then used the median RT for each individual and condition in the group

223 analysis.

224 Pre-planned one-tailed paired-samples t-tests were calculated to assess RT savings

225 (decreases) of the different conditions with respect to baseline. Bonferroni corrections were

226 applied to adjust the p-level for multiple comparisons. Overall differences of combined,

227 timing and order against baseline conditions were corrected for three comparisons (p =

228 puncorrected*3). Early and late test phase differences of these conditions against baseline were

229 corrected for 6 comparisons (p = puncorrected*6), respectively. A repeated-measures ANOVA

230 was calculated in SPSS Statistics (Version 20) to test for an interaction between condition

231 (timing and order transfer) and test phase (early and late).

232 In addition, error rate (% correct trials) in early and late test phases were calculated to

233 evaluate the possibility that RT differences amount to different speed-accuracy strategies.

234 Drift-diffusion model

235 Activity in the five units of the selection layer Z corresponding to fingers was

236 modelled for each stimulus-response event in time steps of 1 ms (as an example only thumb

237 and index finger activity is displayed in Fig. 1). Each unit received input from the visual

238 signal (S), the ordinal (P), and either the temporal (T) or the combined (C) representation,

239 depending on the model. The integration of these inputs determined the rate at which the unit

240 approached the response threshold (100 arbitrary units). When one of the two finger units of

241 the selection layer Z reached threshold, a response was triggered and the model was reset. The

242 corresponding time point n relative to stimulus onset was defined as response time (RT).

243 We compared three models of interaction between the temporal and ordinal

244 representations. Note that the purpose of the model was to illustrate the qualitative differences

245 in the predicted transfer and that these qualitative predictions are stable over a wide range of

246 parameter setting. In the combined model (Fig. 1A) the temporal tuning was linked to a
 11

247 specific ordinal sequence. Each unit in the selection layer was updated as Zn+1 = a * Zn + S +

248 b * P + b * C + ε.

249 In the additive model (Fig. 1B), the temporal representation was separate from the

250 ordinal sequence and each unit was updated as Zn+1 = a * Zn + S + b * P + b * T + ε. Finally,

251 in the multiplicative model (Fig. 1C) each unit was updated as Zn+1 = a * Zn + S + b * P + (P

252 * T) + ε. Note that the ordinal sequence representation (sequence of finger presses) P was

253 added to the activity in the selection layer in all three models based on previous findings that

254 consistently show a representation of movement to be independent of the timing with which it

255 is produced (cf. Introduction).

256 In all three models, the visual stimulus (S) was modelled as a step response, increasing from

257 zero to 1 on the cued finger at time step n=500ms. ε was the Gaussian noise centred at zero

258 with constant variance σ2 = 1. The ordinal representation provided predictive input specific to

259 the anticipated finger, a step response increasing from 0 to 1 right after the completion of the

260 previous response. Therefore, the ordinal input was tonic and did not contain a temporal

261 representation. In contrast, the combined and the temporal representation increased linearly

262 from 0 to 1 starting 100 ms before the stimulus onset to stimulus onset (phasic activation). In

263 the combined representation the output was specific to the upcoming finger in the trained

264 sequence (Fig. 1A). In contrast, in the temporal representation, the output was unspecific to

265 the finger and was sent to all fingers (Fig. 1B and C). The forgetting rate a of Z was set to

266 0.99. The weights of the ordinal, the temporal and the combined representations (b) were set

267 at 0.1. These modulated the rate with which the activity in the selection layer approached the

268 threshold.

269 The predicted RT (time point n at which activity reaches threshold) is plotted in

270 corresponding bar graphs for 100 iterations of each stimulus-response event (Fig. 1A-C, right

271 panel), respectively. We considered the prediction of each model regarding savings in RT
 12

272 relative to a baseline condition (timing and order new) for conditions in which either timing,

273 order or both is preserved. The left part of the bar diagram (early transfer) predicts immediate

274 transfer when the new dimension is introduced. The right part of the bar diagram (late

275 transfer) predicts late transfer, once the new dimension has become more familiar after a few

276 repeated executions of the test sequence.

277 The transfer of timing in the multiplicative model requires short-term learning of the

278 new order to occur after a few repetitions. Previous findings suggest that a sequence of nine

279 key presses can be learned within a couple of trials and faster than the temporal sequence

280 (Ullén and Bengtsson, 2003). In all three models, we therefore assumed that the ordinal

281 representation was learned to 70%, whereas the temporal and combined representations were

282 learned to 10% in the late phase. Since the new ordinal representation is acquired to 70% in

283 the baseline, too, the RT savings of the order condition relative to baseline (new order, new

284 timing) were predicted to decrease in the late transfer phase across the three alternative

285 models.

286

287 Results

288 The RTs of the participants in the training phase of our experiments indicate that

289 subjects learned the task and started to anticipate the correct finger at the correct time. To

290 examine the effect of training, we tested the difference between the first two and the last two

291 blocks of the training phase. In Experiment 1 (Fig. 3A), subjects’ RT significantly decreased

292 from 310 ms (SE = 20) to 251 ms (SE = 11), t(15) = 2.84, p = .006. In Experiment 2 (Fig.

293 3D), RT significantly decreased from 332 ms (SE = 11) to 240 ms (SE = 16), t(14) = 5.14, p <

294 .001. At the end of training a substantial portion of the RTs were faster than 200ms, indicating

295 stimulus anticipation (Hick 1952).

 13

296 To evaluate whether subject acquired sequence specific knowledge through training

297 we compared the RT of the combined (timing and order trained) and the baseline (timing and

298 order not trained) conditions. In Experiment 1 the overall difference between combined (273

299 ms, SE = 10) and baseline (298 ms, SE = 12) was significant, t(15) = -4.512, p < .001,

300 Bonferroni corrected. This result could be replicated in Experiment 2, in which combined

301 (252 ms, SE = 17) had a significantly lower RT than baseline condition (286 ms, SE: 11),

302 t(14) = -3.392, p = .012, Bonferroni corrected (Fig. 3F). This confirmed that subjects acquired

303 knowledge of the trained sequence.

304 All three models predicted that the trained order of finger presses would be transferred

305 to a new timing. In contrast, the models make divergent predictions regarding the transfer of

306 timing. The combined model (Fig. 1A) predicts that timing transfer would not occur, since the

307 temporal representation is bound to the specific sequence of finger presses with which it is

308 learned (combined representation). In contrast, the additive model (Fig. 1B) predicts that the

309 transfer of timing to a new order of finger presses will occur immediately, even though the

310 order is not learned, and would persist in the late phase once the new order is acquired.

311 Finally, the multiplicative model predicts that the timing would not be transferred when the

312 order of finger presses is new, but will emerge once the order of finger presses has become

313 familiar.

314 Immediately at the beginning of each test block (trials 1-3), the timing condition

315 showed only a benefit of 2 ms when compared to baseline, where both timing and order were

316 new, but fixed throughout the respective block. In contrast, the condition with trained order

317 and new temporal structure revealed a benefit of 23 ms compared to baseline (t(14) = -3.155,

318 p = .018, Bonferroni corrected) relative to baseline. However, at the end of each test block

319 (trials 8-10), this relationship reversed: the timing was 29 ms lower than the respective

320 baseline (t(14) = -3.956, p = .006, Bonferroni corrected), whereas order exhibited only a non-
 14

321 significant decrease of 5 ms (Fig. 3B and C). Transfer was computed by subtracting the

322 individual baseline from the timing and order conditions, respectively. A repeated measures

323 ANOVA with the factors transfer (timing transfer, order transfer) by test phase (early: trials 1-

324 3, late: trials 8-10) confirmed a significant interaction between the two factors (F(1,15) =

325 10.14, p = .006, Fig. 3C). This result corroborated our hypothesis that in contrast to finger

326 order the transfer of timing emerges only at a later stage, when the repetition of the new order

327 allows the subjects to recombine the old timing with a new order.

328 These results also indicate that the ordinal sequence of movements is learned more

329 quickly than the temporal sequence – indeed so quickly that it was completely learned within

330 each block of 10 trials. On trial 8-10 of each block, the order condition was not significantly

331 different from baseline (t(15) = -0.517, p = .306, one-tailed, uncorrected), and the temporal

332 condition was not significantly different from the combination condition (t(15) = -0.119, p >

333 .453, one-tailed, uncorrected, Fig. 3C). This fast ordinal learning may be explained by the fact

334 that the stimuli in the current task explicitly instructed the effectors by digits, thereby

335 facilitating a conscious verbal strategy for the ordinal sequence as opposed to the temporal

336 sequence. Indeed, subjects’ reports indicate that they tended to be more aware of the ordinal

337 as opposed to the temporal sequence during the task (Wilcoxon signed-rank test: Z = -1.89; p

338 = .06; mean ordinal awareness: 1.68; and mean temporal awareness: 1.94; on a Likert scale of

339 1 always aware to 5 never aware). Nevertheless, the temporal structure of the sequences was

340 equally task relevant, since subjects did not gain points, if the pressed the corresponding

341 finger to late or too early (cf. Methods).

342 In Experiment 2 we attempted to replicate these findings, while simultaneously

343 contrasting these within the same participants to the condition used by Shin and Ivry (2002)

344 and O’Reilly et al. (2008), in which learned temporal structure was paired with a (pseudo-)

345 random sequence of finger presses. To replicate these results in our paradigm, we tested for
 15

346 transfer of learned timing to a pseudo-random order. We compared a baseline condition in

347 which both timing and order were random on each trial (baselineR) to a condition with trained

348 timing and random order on each trial (timingR). We also replicated the baseline condition, in

349 which both timing and order was new, but fixed across the test block, and the timing condition

350 with trained timing and fixed new order from Experiment 1.

351 As can be seen from Figure 3F, the overall RTs in the timing condition were

352 significantly different from baseline (t(14) = -2.592, p = .033, Bonferroni corrected), with a

353 decrease in mean RT emerging in the second part of each test block (Fig. 3E). This replicated

354 our results from Experiment 1. In contrast, there was no significant difference between the

355 timingR and baselineR conditions (t(14) = -0.64, p = .798, Bonferroni corrected), replicating

356 the lack of significant RT costs associated with a random probe in the timing conditions (Shin

357 and Ivry, 2002; O’Reilly et al., 2008). Importantly, the transfer of the learned temporal

358 structure in the timing condition (22 ms, SE = 8) was significantly stronger than the transfer in

359 the timingR condition (3 ms, SE = 4), t(14) = -2.603, p = .01 (Fig. 3F). This finding again

360 shows that an independent representation of timing can only be revealed when the trained

361 temporal sequence is recombined with a new, stable, ordinal sequence.

362 Discussion

363 To address divergent findings with regard to the representation of temporal structure in

364 movement sequences, we hypothesized three alternative models regarding the representation

365 of timing and the interaction of timing and order. The combined model hypothesized that

366 temporal structure is bound to the ordinal structure of movements (finger sequence) with

367 which it was acquired and when decoupled from this ordinal structure shows no savings

368 relative to an untrained sequence. This model would be in accordance with previous SRTT

369 studies (Shin and Ivry, 2002; O’Reilly et al., 2008). In contrast, the additive model

370 hypothesized that the temporal structure is represented independently of the ordinal structure
 16

371 of finger movements, increasing all eligible movement signals at specific points in time.

372 Accordingly, a known temporal structure is expected to produce performance advantages

373 even when paired with a new sequence of finger movements. Finally, a multiplicative model

374 hypothesized that the nervous system acquires an effector-independent representation of

375 temporal structure, but in contrast to the additive model its signals interact multiplicatively

376 with the signals originating from the ordinal representation. Accordingly, timing signals from

377 the temporal representation modulate only the expected ordinal structure and performance

378 advantages of a trained temporal structure can only occur once the new ordinal structure

379 becomes familiar, but not when it is unknown.

380 We conducted two experiments in which we used a temporally structured serial

381 reaction time task with visually cued responses. Our results offer strong support for the

382 multiplicative model (Fig. 1C and Fig. 3). Towards the end of each test block (late transfer

383 phase) that consisted of 10 sequence repetitions, shorter reaction times (RT) occurred when

384 subjects produced a new order of finger movements with a timing that was previously trained,

385 as compared to a baseline in which both timing and order were new (Experiment 1 and 2). In

386 contrast, when the order changed on each trial and test block, the previously trained timing

387 did not show any performance advantages (Experiment 2). The latent temporal representation

388 could be revealed because of the fast learning rate for the ordinal representation, which is in

389 line with previous findings demonstrating that subjects learn order faster than timing (Ullén

390 and Bengtsson, 2003).

391 The RT findings did not amount to a change in strategy such as slowing down

392 responses to be more accurate in the early phase or speeding up in the late phase while

393 sacrificing fidelity in finger order of each timing test block. On the contrary, error rate

394 paralleled the RT results (Fig. 4): Subjects made more errors in the early as opposed to the
 17

395 late phase of each timing test block in both experiments. This suggests that the reported RT

396 findings reflect a true temporal transfer effect.

397 Our findings reconcile contradictory results from previous studies (Salidis, 2001; Shin

398 and Ivry, 2002; Ullén and Bengtsson, 2003; O’Reilly et al., 2008; Gobel et al., 2011)

399 regarding the representation of timing in sequential finger movements. O’Reilly et al. (2008)

400 specifically showed that the cost of changing the order was as high as changing both order

401 and timing. We confirmed this result, but could show that performance advantages emerge

402 compared to baseline (where both timing and order changed) once subjects become familiar

403 with the sequence of movements they have to perform. This demonstrates that the nervous

404 system can acquire a temporal representation in a reaction time task and transfer it across

405 different sequences of muscle commands, unlike previously suggested (Shin and Ivry, 2002;

406 O’Reilly et al., 2008). Notably, the above studies have also demonstrated that temporal

407 learning does not occur unless timing is associated with a specific order in the first place.

408 Indeed, the transfer in our study occurred after training of a temporal sequence in combination

409 with a fixed finger order. Taken together, this suggests that the transferable temporal

410 representation can only be acquired when initially linked to a specific movement order.

411 The current design differs from the SRTT in previous studies (Salidis, 2001; Shin and

412 Ivry, 2002; O’Reilly et al., 2008) insofar as it employs shorter sequences and only one cycle

413 of each sequence per trial. It is likely that this design led to more awareness of the ordinal and

414 temporal structure than traditional SRTT designs. While our post-experimental questionnaire

415 rating indeed suggests awareness of the ordinal and temporal sequential structure, within the

416 same paradigm we were able to replicate both transfer and failure of transfer of temporal

417 structure, as found with the more traditional SRTT tasks (Shin and Ivry, 2002; O’Reilly et al.,

418 2008). We therefore expect that our findings will generalize to more implicit versions of the

419 task.
 18

420 Our model assumes that with training of a finger sequence an ordinal representation of

421 the sequence emerges, which tonically pre-activates the next movement in the selection layer

422 prior to the imperative cue. Electrophysiological measurements in the (pre-) supplementary

423 motor area (SMA) and the premotor cortex in monkeys have shown that after movement

424 sequence training different neurons can become tuned to movement transitions or whole

425 sequences (Mushiake et al., 1991; Shima and Tanji, 2000). Activity in these neurons is up-

426 regulated during each interval between movements of a specific movement transition or

427 preceding a specific movement sequence, respectively. This tonic preparation is in line with

428 the neuronal activity increase reported in direction-selective premotor (but not primary motor)

429 cortex during preparatory movement selection (Cisek and Kalaska, 2005). Since the

430 foreperiod varied randomly across trials (within a certain time range) in this study, this

431 increase is likely to be driven by movement preparation alone.

432 Our results support the hypothesis that the nervous system also acquires an

433 independent temporal representation of the trained sequence, which phasically boosts the

434 tonic signals emitted from the ordinal representation at the predicted time of the visual cue.

435 Neuroimaging studies have reported regions that may be associated with temporal sequence

436 representation in movements (Ramnani and Passingham, 2001; Bengtsson et al., 2004), i.a.

437 premotor areas, superior temporal gyrus and parts of the cerebellum. While there is no

438 neurophysiological data on the production of timed finger sequences, direct stimulation of the

439 circuitries enabling eye-blink conditioning tasks suggests that individual Purkinje cells can be

440 trained to respond with temporal profiles adapted to a specific inter-stimulus-interval between

441 conditioned responses (Jirenhed and Hesslow, 2011a, 2011b). Towards the end of

442 conditioning, Purkinje cells exhibit a decrease of simple spike activity in anticipation of the

443 air-puff leading to an accurately timed response. However, it is not known whether Purkinje

444 cells, similarly to cortical cells in the pre-SMA/SMA, can also code for certain transitions or
 19

445 whole sequences of temporal intervals or whether interval specific coding can be utilized

446 across effectors.

447 In our model, the temporal representation was specified as a phasic increase of activity

448 briefly before the anticipated onset of the cue modulating the drift rate of the selection layer.

449 Alternative ways may be employed to model the temporal representation. Instead of phasic

450 increases suggested in our model, temporal information could be an amplitude-modulated

451 version of the tonic ordinal representation with higher amplitude for shorter intervals between

452 two movements. Crucially, however, adjusting the drift rate based on the position in the

453 sequence would also imply that the nervous system has a representation of the temporal

454 structure.

455 The multiplicative model can also be applied to studies of choice reaction time tasks.

456 Of particular interest here are TMS studies. TMS can be taken to probe the excitability of

457 units in the selection layer, since TMS pulses are delivered over motor regions with

458 connections to the periphery via the corticospinal tract. Mars et al. (2007) and Van Den Hurk

459 et al. (2007) demonstrated that knowledge about the effector can decrease RT and increase

460 cortico-spinal excitability of the respective effector in the preparation phase (foreperiod).

461 Since the duration of the foreperiod was random and the preparatory increase in excitability

462 prolonged, this effect reflects effector-specific preparation rather than temporal preparation,

463 although the likelihood of target occurrence (hazard rate of elapsed time) may also play a role

464 (Nobre et al., 2007).

465 In contrast, when the timing is known due to a fixed foreperiod, suppression of

466 cortico-spinal excitability has been reported specifically for the precued effector (Touge et al.,

467 1998; Davranche et al., 2007; Duque and Ivry, 2009; Duque et al., 2012). The suppression

468 becomes more prominent with proximity to the expected imperative cue, then turning into

469 selective facilitation at imperative cue onset. Such a time dependent modulation has been
 20

470 associated with faster RT and it is likely to be related to temporal preparation to ensure that

471 the response does occur at the correct point in time. Although our model does not assume that

472 the temporal representation can provide inhibitory input to the selection layer, our hypothesis

473 of a multiplicative interaction between temporal and ordinal signals is congruent with the

474 finding that the inhibitory effects are especially pronounced for the precued effector.

475 Our findings did fail to show immediate transfer (early test phase) of the temporal

476 sequence when a novel finger movement sequence is cued. This result appears at odds with

477 the foreperiod effect observed in a choice-RT task (Bertelson and Boons, 1960). This study

478 has shown reaction times for fixed foreperiods even though the cued effector is not known

479 during the foreperiod. As only two fingers have been task-relevant in the latter study, it is

480 possible that both are activated to an equal extent and modulated similarly by the temporal

481 signal in the expected fixed period condition. Both additive and a multiplicative integration of

482 effector and timing interaction could explain this behaviour and future studies should address

483 whether the number of relevant effectors, as well as their sequential organisation in time

484 modulate the interaction between temporal and effector representations.

485 Non-linear multiplicative interaction between different representation is a common

486 computational tool in the nervous system, such as during sensorimotor transformations in

487 parietal neurons (Andersen et al., 1985; Pouget and Sejnowski, 1997), auditory space

488 localization in the subcortical auditory pathway (Peña and Konishi, 2004), or the integration

489 of multisensory input (Huston and Krapp, 2009). On the level of a single neuron,

490 multiplicative neural interaction may be achieved by a clustering of synapses at the dendritic

491 tree (Koch and Segev, 2000). Accordingly, if synapses cluster together on neighbouring

492 dendritic patches they can cooperate in activating the local excitable dendritic channels,

493 thereby elevating the firing probability of the respective neuron. Alternatively, a

494 multiplicative interaction of timing and finger order could arise from neural units with linear
 21

495 summation of the input combined with an output non-linearity (Dayan and Abbott, 2001).

496 Future studies should address how the nervous system represents and integrates temporal and

497 ordinal sequential information for skilled movement.

498 In conclusion, our behavioural results reveal that sequential movement timing is

499 represented independently of the sequence of muscle commands, but modulates the motor

500 output in a non-linear multiplicative fashion.

501

502 Acknowledgements

503 We would like to thank Richard Ivry, Alexandra Reichenbach and Naveed Ejaz for

504 comments on an earlier draft of this manuscript.

505 Grants

506 This work was supported by the Marie Curie Initial Training Network “Cerebellar-

507 cortical control: Cells, circuits, computation, and clinic“.

508

509 Author contributions

510 KK and JD conceived and designed the study; KK and AS preformed experiments and

511 analyzed data; KK, AS and JD interpreted the results; KK, AS and JD wrote the manuscript.

512

513

514 References

515 Andersen RA, Essick GK, Siegel RM. Encoding of spatial location by posterior parietal
516 neurons. Science 230: 456–458, 1985.

517 Bengtsson SL, Ehrsson HH, Forssberg H, Ullén F. Dissociating brain regions controlling
518 the temporal and ordinal structure of learned movement sequences. Eur. J. Neurosci. 19:
519 2591–2602, 2004.
 22

520 Bertelson P, Boons J-P. Time Uncertainty and Choice Reaction Time. Nature 187: 531–532,
521 1960.

522 Cisek P, Kalaska JF. Neural correlates of reaching decisions in dorsal premotor cortex:
523 specification of multiple direction choices and final selection of action. Neuron 45: 801–814,
524 2005.

525 Conditt MA, Mussa-Ivaldi FA. Central representation of time during motor learning.
526 Proceedings of the National Academy of Sciences 96: 11625 –11630, 1999.

527 Davranche K, Tandonnet C, Burle B, Meynier C, Vidal F, Hasbroucq T. The dual nature
528 of time preparation: neural activation and suppression revealed by transcranial magnetic
529 stimulation of the motor cortex. European Journal of Neuroscience 25: 3766–3774, 2007.

530 Dayan P, Abbott LF. Theoretical Neuroscience: Computational and Mathematical Modeling
531 of Neural Systems. 1st ed. The MIT Press, 2001.

532 Van Den Hurk P, Mars RB, Van Elswijk G, Hegeman J, Pasman JW, Bloem BR, Toni I.
533 Online Maintenance of Sensory and Motor Representations: Effects on Corticospinal
534 Excitability. J Neurophysiol 97: 1642–1648, 2007.

535 Diedrichsen J, Shadmehr R, Ivry RB. The coordination of movement: optimal feedback
536 control and beyond. Trends Cogn. Sci. (Regul. Ed.) 14: 31–39, 2010.

537 Duque J, Ivry RB. Role of Corticospinal Suppression during Motor Preparation. Cerebral
538 Cortex 19: 2013 –2024, 2009.

539 Duque J, Labruna L, Verset S, Olivier E, Ivry RB. Dissociating the Role of Prefrontal and
540 Premotor Cortices in Controlling Inhibitory Mechanisms During Motor Preparation. J.
541 Neurosci. 32: 806–816, 2012.

542 Gobel EW, Sanchez DJ, Reber PJ. Integration of temporal and ordinal information during
543 serial interception sequence learning. Journal of Experimental Psychology: Learning,
544 Memory, and Cognition 37: 994–1000, 2011.

545 Hick WE. On the rate of gain of information. Quarterly Journal of Experimental Psychology
546 4: 11–26, 1952.

547 Huston SJ, Krapp HG. Nonlinear Integration of Visual and Haltere Inputs in Fly Neck
548 Motor Neurons. J. Neurosci. 29: 13097–13105, 2009.

549 Jirenhed D-A, Hesslow G. Learning stimulus intervals--adaptive timing of conditioned

550 purkinje cell responses. Cerebellum 10: 523–535, 2011a.

551 Jirenhed D-A, Hesslow G. Time Course of Classically Conditioned Purkinje Cell Response
552 Is Determined by Initial Part of Conditioned Stimulus. The Journal of Neuroscience 31: 9070
553 –9074, 2011b.

554 Koch C, Segev I. The role of single neurons in information processing. Nat. Neurosci. 3
555 Suppl: 1171–1177, 2000.
 23

556 Mars RB, Bestmann S, Rothwell JC, Haggard P. Effects of motor preparation and spatial
557 attention on corticospinal excitability in a delayed-response paradigm. Exp Brain Res 182:
558 125–129, 2007.

559 Mushiake H, Inase M, Tanji J. Neuronal activity in the primate premotor, supplementary,
560 and precentral motor cortex during visually guided and internally determined sequential
561 movements. J. Neurophysiol. 66: 705–718, 1991.

562 Nissen MJ, Bullemer P. Attentional requirements of learning: Evidence from performance
563 measures. Cognitive Psychology 19: 1–32, 1987.

564 Nobre A, Correa A, Coull J. The hazards of time. Curr. Opin. Neurobiol. 17: 465–470,
565 2007.

566 O’Reilly JX, McCarthy KJ, Capizzi M, Nobre AC. Acquisition of the temporal and ordinal
567 structure of movement sequences in incidental learning. J. Neurophysiol. 99: 2731–2735,
568 2008.

569 Oldfield RC. The assessment and analysis of handedness: the Edinburgh inventory.
570 Neuropsychologia 9: 97–113, 1971.

571 Peña JL, Konishi M. Robustness of multiplicative processes in auditory spatial tuning. J.
572 Neurosci. 24: 8907–8910, 2004.

573 Penhune VB, Steele CJ. Parallel contributions of cerebellar, striatal and M1 mechanisms to
574 motor sequence learning. Behav. Brain Res. 226: 579–591, 2012.

575 Pouget A, Sejnowski TJ. Spatial Transformations in the Parietal Cortex Using Basis
576 Functions. Journal of Cognitive Neuroscience 9: 222–237, 1997.

577 Ramnani N, Passingham RE. Changes in the Human Brain during Rhythm Learning.
578 Journal of Cognitive Neuroscience 13: 952–966, 2001.

579 Ratcliff R, McKoon G. The Diffusion Decision Model: Theory and Data for Two-Choice
580 Decision Tasks. Neural Computation 20: 873–922, 2007.

581 Ratcliff R. A theory of memory retrieval. Psychological Review 85: 59–108, 1978.

582 Robertson EM. The Serial Reaction Time Task: Implicit Motor Skill Learning? J. Neurosci.
583 27: 10073–10075, 2007.

584 Salidis J. Nonconscious temporal cognition: Learning rhythms implicitly. Memory &
585 Cognition 29: 1111–1119, 2001.

586 Shima K, Tanji J. Neuronal activity in the supplementary and presupplementary motor areas
587 for temporal organization of multiple movements. J. Neurophysiol 84: 2148–2160, 2000.

588 Shin JC, Ivry RB. Concurrent learning of temporal and spatial sequences. J Exp Psychol
589 Learn Mem Cogn 28: 445–457, 2002.

590 Shin JC, Ivry RB. Spatial and Temporal Sequence Learning in Patients with Parkinson’s
591 Disease or Cerebellar Lesions. Journal of Cognitive Neuroscience 15: 1232–1243, 2003.
 24

592 Touge T, Taylor JL, Rothwell JC. Reduced excitability of the cortico-spinal system during
593 the warning period of a reaction time task. Electroencephalography and Clinical
594 Neurophysiology/Electromyography and Motor Control 109: 489–495, 1998.

595 Ullén F, Bengtsson SL. Independent processing of the temporal and ordinal structure of
596 movement sequences. J. Neurophysiol. 90: 3725–3735, 2003.

597 Xuan B, Chen X-C, He S, Zhang D-R. Numerical magnitude modulates temporal
598 comparison: An ERP study. Brain Research 1269: 135–142, 2009.

599

600

601
 25

602 Figure legends

603 Figure 1: Alternative hypotheses of the interaction between ordinal and temporal

604 representations in a timed serial reaction time task and its implications for transfer of the

605 trained temporal sequence. At the beginning of learning, the motor commands are exclusively

606 triggered by a sequence of cues (illustrated for index followed by thumb) through a learned

607 stimulus-response association S (imperative cue – finger mapping). This layer triggers activity

608 in the corresponding unit of the selection layer Z, which in turn issues a correct response when

609 the activation reaches threshold. With learning, the motor system acquires a representation of

610 the ordinal sequence of finger movements P that predicts the next response in form of a tonic

611 pre-activation of the effector, which is sent to the corresponding unit in the selection layer.

612 (A) Combined model: Furthermore, the nervous system is hypothesized to acquire a combined

613 sequence representation C, which contains a more precisely timed ordinal representation, that

614 phasically increases the predicted unit just before the expected times of cue onset (left panel).

615 Since the timing of movements is bound to the learned order of movements, no transfer (RT

616 advantage to baseline) of temporal structure can occur when the latter is tested with a new

617 ordinal structure (right panel). (B) Additive model: An independent representation of the

618 temporal sequence is acquired that sends phasic signals to all the units just before the

619 expected times of cue onset (left panel). Temporal structure would be transferrable to a new

620 order of movements (right panel) even when the ordinal structure with which it is recombined

621 is completely unknown. (C) Multiplicative model: An independent representation of temporal

622 structure is acquired, that is integrated with the ordinal signal multiplicatively, before it is

623 propagated to the selection layer. Transfer can only occur once the ordinal structure with

624 which it is paired is partly learned (right panel, late transfer phase), because the temporal

625 representation can only take effect on pre-activated units.

 26

626 Figure 2: Task design. (A) Subjects performed a SRTT by producing a force press

627 with the selected finger of the right hand as fast and accurately as possible after a visual

628 presentation of an imperative stimulus (e.g. “4” ring finger). The cues were separated by a

629 specific sequence of inter-stimulus-intervals. Reaction time was defined as force exceeding

630 0.6 N both for online feedback and data analysis. (B) After training (200 trials) on the

631 combination of a specific sequence of finger presses and temporal intervals, participants were

632 tested on different condition in blocks of 10 trials each. In Experiment 1, subjects were tested

633 on the trained sequence (“combined”), on the trained temporal sequence combined with a new

634 ordinal sequence (“timing”), on the trained ordinal sequence combined with a new temporal

635 sequence (“order”), as well as on new ordinal and temporal sequence (“baseline”). In all these

636 conditions, the sequence stayed the same for a block of 10 trials. In Experiment 2 subjects

637 were also tested on the trained temporal sequence combined with a random finger order

638 (“timingR”) and on a random temporal sequence combined with a random ordinal sequence

639 (“baselineR”). The order condition was not tested in the latter.

640 Figure 3: Reaction time (RT) results: Reaction time decreased from the beginning to

641 the end of training by around 100 ms in both experiments (A and D). Within each test block

642 RT decreased when order and timing was repeated across trials (B and E), but remained on

643 the same level when order or both timing and order were random (E). A significant decrease

644 from baseline appeared in the timing condition, in particular in the late phase, after the test

645 sequence has been repeated across each block (C and F). This decrease could not be observed

646 when finger order changed on every trial (F). Error bars denote standard error of the mean.

647 Figure 4: Error rate (% correct trials) in early and late phases of Experiments 1 (A)

648 and 2 (B) across subjects paralleled RT results. The error rate significantly decreased in the

649 late compared to the early phases of Experiment 1 (t(15) = 2.905, p = 0.005) and Experiment

650 2 (t(14) = 2.092, p = 0.028). This suggests that the reported RT findings do not amount to a
 27

651 change in strategy – slowing down responses to be more accurate in the early phase or

652 speeding up in the late phase sacrificing fidelity in finger order – but reflect a learning-related

653 transfer effect.

654

655

656