Neural variability in premotor cortex provides a signature of motor preparation

Mark M Churchland, Byron M Yu, Stephen I Ryu, Gopal Santhanam, Afsheen Afshar, Krishna V Shenoy

How does the brain generate rapid and accurate movements of the body? Much

work has focused on the role of sensory feedback and internal models in optimizing control signals during movement. Here we explore the possibility that the optimization of control signals begins during motor preparation, before movement begins. Using a delayed reach paradigm, delay-period or ‘preparatory’ activity can be observed in primary motor cortex (M1), and is even more prevalent in dorsal premotor cortex (PMd). It is often suggested that such activity is related to motor programming. This hypothesis implies that a movement is in large part the product of a motor ‘program’ latent in the preparatory activity. If so, it would seem critical that the brain optimize preparatory activity, so as to get the desired result when the movement is triggered. Given the presumably non-linear mapping from preparatory activity to movement (and thence to likelihood of reward) one suspects that this optimization is reasonably difficult, and might consume considerable time.

We present evidence, from three sets of experiments, that such optimization does indeed occur during motor preparation. Recordings were made from PMd of two rhesus monkeys. In the first set of experiments, we asked whether trial-by-trial variability in the state of preparatory activity correlates with trial-by-trial variability in the actual movement. Movements made by biological organisms are inherently variable, and this is usually thought to result from noise sources that contribute during the movement. In contrast, we show that a considerable portion of this variability is due to variability in preparatory activity. In particular, delay-period activity was predictive of the natural variability in the peak speed of the upcoming movement (p<0.001 for both monkeys). These data indicate that the state of preparatory activity has consequences for the upcoming movement.

In the second set of experiments, we examined the across-trial variability of firing rates in PMd. Our hypothesis was that, before target onset (before optimization begins) firing rates would likely be variable. However, after optimization is complete, firing rates would be more consistent, assuming that the range of ‘optimal’ states is limited. We did indeed find that the across-trial variability of PMd firing rates dropped following target onset, with a time-course of 100-200 ms. Variability was assessed by computing the variance, and normalizing by mean firing rate, to exclude the influence of Poisson spiking variability. The time-course of this normalized variance (NV) measurement was found to be consistent with the time-course of motor preparation, inferred behaviorally from reaction time (RT) measurements. We repeated the experiment using three discrete delay period durations (30,130,230 ms). Across these conditions, the NV, measured at the time of the go cue, was predictive of the subsequent RT. When we examined the

Computational and Neural Systems (COSYNE) Annual Meeting, Slide Presentation, 2005. In Press.Published abstract (pages 1-2) and supplementary material for COSYNE reviewers.

natural distribution of RT’s, we found that trials where the NV was lower had shorter RT’s. It thus appears that the NV provides a rough signature of the progress of motor preparation. Specifically, we suggest that the NV captures the process by which PMd firing rates travel from a variable initial state to a more consistent, and presumably near-optimal, final state. This application of the NV measurement (which is similar to the Fano-factor) provides a new way of interrogating the function of recorded neurons. Rather than asking under which conditions neurons maximally change their firing rates, the NV asks under which conditions their firing rates become consistent.

The above results suggest that the brain is capable of monitoring the state of motor preparation, and delaying movement if further optimization is necessary. If so, then disrupting preparatory activity ought to delay RT. In the third set of experiments, we used microstimulation to disrupt activity in PMd and M1, either during the delay period or just after the go cue. Microstimulation of PMd just after the go cue caused an increase in mean RT (two monkeys; 30 and 73 sites, mean increase of 26 and 20 ms, p<10-6 for each). In constrast, microstimulation of M1 had almost no effect on RT (12 and 32 sites, mean increase of 4 and 2 ms, p>0.1). Our interpretation is that when motor preparation is disrupted just before it is to be translated into movement, execution is delayed until preparatory activity can be re-optimized. The discrepancy between M1 and PMd is consistent with the greater prevalence of delay-period activity in PMd. Microstimulation appears to selectively disrupt motor preparation related to the reach: there was only a weak effect on saccadic RT (8 and 2 ms, p<0.001 and p>0.1). Microstimulation delivered 100 ms or more before the go cue had only a slight effect on RT (7 and 2 ms), consistent with estimates of the time-course of motor preparation provided by the NV analysis above. Lastly, microstimulation just before the go cue delayed RT, but only if the delay-period was long enough for motor preparation to have been at least partially accomplished (ANOVA interaction, p<0.05 for each).

The above results provide converging evidence that PMd preparatory activity is optimized prior to movement generation. This process appears to take time, possibly explaining why RT’s of voluntary movements are longer than expected from conduction delays alone. Optimization also appears to take variable time, possibly explaining why RT’s are naturally variable. The end result of this process also appear to be variable, revealing at least one source of movement variability.

Our results are consistent with prior suggestions that a ‘forward model’ could be used by the brain to optimize motor preparation. In this conception, the forward model would act in concert with a controller, as part of a recurrent circuit that would act to optimize activity. In general, artificial recurrent networks are capable of performing optimizations – typically the optimal solution forms an attractor in the dynamics of the networks. We note that a common property of such models is that activity is variable before optimization is complete, and consistent afterwards, in keeping with our observations that the across-trial variability of PMd neurons drops during motor preparation. Support: Helen Hay Whitney, NIH, BWF, NSF, ONR, Sloan, Whitaker.

Computational and Neural Systems (COSYNE) Annual Meeting, Slide Presentation, 2005. In Press.Published abstract (pages 1-2) and supplementary material for COSYNE reviewers.

Neural variability and motor preparation Churchland et al.

Motor preparation is often studied using the ‘instructed delay’ paradigm, in which an instruction stimulus is temporally separated from a go cue(1-6). For example, the onset of a reach target might be followed by a variable delay period, after which a go cue would signal that the reach should be executed. Longer delay periods typically lead to shorter reaction times (RT’s), and this has been interpreted as evidence for a motor preparation process that takes time(2, 5-7). According to this view, delay periods allow motor preparation to complete before the go cue, shortening RT. Presuming this interpretation is correct, why does motor preparation consume time? Does it involve primarily ‘high level’ processes: the shifting of attention, intent, or even the eyes themselves? Or does it involve detailed motor programming? Most importantly, what sort of process might it be at the neural level?

Our hypothesis has been that motor preparation involves the time-consuming optimization of preparatory activity in the motor cortices, especially in dorsal premotor cortex (PMd), which shows a prevalence of delay-period activity. Central to this view is the idea that a movement is in large part a product of the preparatory activity that precedes it. It then follows that the brain must optimize such activity, so as to produce a movement very similar to that desired. Of course, it may be that purely feedforward connections (e.g. from paretial cortex) are sufficient to create optimal preparatory activity in a single step. However, it seems likely that feedforward drive is not sufficient, and that recurrent connections act to refine preparatory activity over time(8). Such connections exist both within the motor cortices, and between them and the basal ganglia and cerebellum(9, 10). Here we present three experimental tests of this hypothesis. Experiment 1. We asked whether the executed movement depends upon the precise state of preparatory activity. Neural activity was recorded from PMd of two rhesus monkeys who performed a delayed-reach task. Monkeys were trained to reach at either a fast or moderate speed depending on the color of the target (red or green). This allowed us to ascertain the speed preference of each neuron. Importantly, there was within-category variability in peak reach speed (e.g. some reaches to red targets were slightly faster than others). This variability is unlikely to be due to incomplete training, or general task sloppiness – monkeys were highly-trained and better at the task that most humans. The variability most likely reflects the fact that there is a limit to how consistent a movement can be(11). We asked whether these trial-by-trial variations in peak speed correlated with trial-by-trial variations in delay-period activity. As illustrated by the examples in figure 1, such correlations were common, and usually agreed with the speed preference of the neuron. This was a statistically significant effect for many individual sites, and across the population for both monkeys (p<0.001 for both).

Figure 1. Trial-by-trial scatterplot of firing rate versus peak reach speed. Each point represents one trial. Cell #29 preferred reaches to the red (fast) target, and data shown are for this target. The slope is positive as predicted. Cell #16 preferred reaches to the green (slow) target, and data shown are for this target. The slope is negative as predicted.

These results demonstrate that the intrinsic variability of voluntary movements is due in part to variability in motor preparation, and is not due solely to online sources of noise(11), as has previously been argued(12). Of course, we cannot conclude with certainty that it is PMd activity itself that actually drives the variability. The evidence implicating PMd is purely correlative – preparatory activity elsewhere could influence both PMd activity and the eventual movement. Still, these data support our hypothesis that the state of preparatory activity has consequences for the upcoming movement. Experiment 2. Figure 2 illustrates a simple view of how the optimization of preparatory activity might occur. For each possible movement there is an ideal configuration of firing rates (or, as illustrated, a subspace of ideal configurations). During motor preparation, firing rates are brought from their initial state to the subspace corresponding to the desired movement. We posit that the brain waits to trigger the movement until after firing rates enter the appropriate subspace. The first prediction of this ‘optimal subspace’ is illustrated in figure 3a. If we consider activity at the time of the go cue, then for some trials the configuration of firing rates will lie within the ideal subspace (motor preparation is complete) and RT’s will

1

Computational and Neural Systems (COSYNE) Annual Meeting, Slide Presentation, 2005. In Press.Published abstract (pages 1-2) and supplementary material for COSYNE reviewers.

Neural variability and motor preparation Churchland et al.

neuron 1

neuron 2

neuron 3

trial 1

trial 2

left reachright reach

firing rate,

be short. For other trials, the configuration of firing rates will lie outsideal subspace (motor preparation is incomplete), and RT’s will be lodifficult to directly measure the location of the ideal subspace; its very existence is hypothetical. Still, we can exploit the fact that the variability of firing rates, measured across trials, is expected to be greater for the long RT trials.

To test this prediction, we made extracellular re

ide the nger. It is

cordings from neurons in PMd

-een

e

of two monkeys. We computed the variance of firing rate, across trials,normalized by mean firing rate. This normalized variance (NV) metric was computed as a function of time for each isolation and target location. For statistical power, measurements of the NV were then averaged across all isolations/locations. The NV is essentially a rate-based (rather than countbased) version of the Fano-factor. However, the Fano-factor has typically bused to gauge intrinsic spiking variability(13-16); additional sources of variability are excluded if possible. Here, any additional variability is thquantity of interest.

Figure 2. Illustration of the optimal subspace hypothesis

Figure 3b plots the NV, around the time of the go cue, for trials with RT’s longer (red) and shorter (green) than the median. Data were collected over 7 days using an electrode array implanted in PMd. Consistent with the above prediction, trials with short RT’s had less variability in firing rate around the time of the go cue. The black trace at the bottom plots the mean percent difference in the NV (short – long RT’s). This was computed by taking the percent difference between the NV for the short and long RT trials for each isolation/condition, and then computing the mean and standard error.

target100 ms

go

1.0

1.4

5

-5

long RT's

short RT's

ba

NV

% d

iffer

ence

short RT's

long RT's

Figure 3. a. Illustration of how RT might be related to firing rate. The shaded area represents the optimal subspace for the movement being prepared. Each dot corresponds to one trial, and represents the configuration of firing rates at the time of the go cue. For some trials, that configuration may lie within the optimal subspace (green dots). For other trials, it may lie outside (red dots). Trials with firing rates within the subspace are expected to have shorter RT’s, as no further changes in firing rate are necessary before the movement can be triggered. b. The NV is plotted as a function of time around the onset of the go cue. Red and green traces show the NV for trials with RT’s longer and shorter than the median. Traces at bottom show the mean percent difference (long minus short) in the NV (black) and mean firing rate (blue). Flanking traces show standard error. Data are from monkey G, and include only trials with delay periods >200 ms. For statistical power, data were pooled across 7 day’s recording.

Similar results were obtained when the analysis was applied to single electrode recordings from a second monkey (B, 51 single-unit isolations). For both monkeys, there was less across-trial firing rate variability for short RT trials, both during the delay period and after the go cue (p<0.05 for all). In contrast, we found an inconsistent relationship between the magnitude of mean firing rate and RT. The blue trace in figure 3b plots the mean percent different in firing rate for trials with RT’s shorter and longer than the median. The difference is negative, indicating that RT’s were shorter when firing rates were lower. This was a statistically significant effect when data were collapsed across the delay period. However, for monkey B, the opposite effect was found. For short RT’s, firing rates were on average 1.4% higher during the delay period (p>0.05). These small and inconsistent differences in mean firing rate suggest that the cartoon in figure 3b is approximately correct, that the mean of the distribution is similar for short and long RT trials, with the primary difference being the distance from that mean.

A second prediction of the ideal subspace hypothesis is illustrated in figure 4a. Before target onset, the monkey need not prepare anything in particular, and activity related to motor preparation might be quite variable. Once motor preparation is complete, activity should occupy the optimal subspace, and therefore be fairly consistent. Given these assumptions, it might be possible to track the time-course of motor preparation

2

Computational and Neural Systems (COSYNE) Annual Meeting, Slide Presentation, 2005. In Press.Published abstract (pages 1-2) and supplementary material for COSYNE reviewers.

Neural variability and motor preparation Churchland et al.

a

200 ms

0

80

1

1.5

rate

NV

trial

s

target

bby tracking the rate at which variability declines. As above, we measure firing rate variability using the NV. The simulations in figure 4b illustrate the behavior of this metric, and assume spike generation with Poisson statistics. If each trial is simulated to have the same underlying firing rate (solid trace at top), the NV (dashed trace at bottom) is constant at unity throughout. If the underlying firing rate on each trial is initially variable (grey traces), the NV (solid trace at bottom) is at first elevated, and drops to unity as firing rates converge.

Figure 5 shows the NV computed from 47 isolations recorded simultaneously from PMd of monkey G during the delayed reach task. The black trace plots the mean NV across all isolations and target locations. The NV declined following target onset, and again following the go cue. From 200 ms before target onset to the median time of the go cue, the NV declined 24% (t-test, p<10-6). By movement onset it had declined 38% (p<10-10). The initial decline in the NV spanned ~119 ms. Data shown are for a single day’s recording. Data from other days are very similar. Similar data were also obtained from two other monkeys (drops in the NV of 19 and 20% by the time of the go cue; 36 and 37% by the time of movement onset; p<0.001 for all; duration of the initial decline ~145 and 198 ms).

Figure 4. a. Illustration of how firing rate variability might be reduced over time. Before motor preparation begins, firing rates are near baseline and variable (lower left cloud of dots). After motor preparation is complete, firing rates lie within the optimal subspace, and are less variable. b. Simulations showing the behavior of the NV. The solid black trace plots the mean firing rate of a single PMd neuron, which forms the basis of the simulations. For each of 10,000 simulated trials (10 of which are shown in the rasters) spike trains were generated using Poisson statistics. In the bottom panel, the dashed trace shows the NV computed when the underlying firing rate of every simulated trial was equal to the mean firing rate in the top panel. The solid trace shows the NV computed when underlying firing rates were initially variable, and converged to the mean. Ten such trials are shown by the grey traces in the top panel.

200 mstargetgo

move

hand speedNV

1

1.5

0

50

cm/s

Figure 5. The NV as a function of time. The black trace plots the mean NV (±SEM) across all isolations and target conditions (both preferred and non-preferred). The grey trace shows mean absolute hand speed, also averaged across all target locations. Two temporal epochs are shown, aligned to target and movement onset, the times of which are indicated by the black arrows. The small solid histogram shows the distribution of go-cue onset times, reflecting the fact that RT’s are variable. (816 trials, 47 isolations: 14 single- and 33 multi-unit).

The above results demonstrate that target onset drives a decline in the variability of PMd neurons. Might there be a trivial explanation for this result? There are a number of possibilities to consider. Might the decline be due indirectly to saccadic behavior, or to small movements of the hand? Might it be due to a change in intrinsic (rather than across-trial) spiking variability – perhaps spiking becomes more regular at higher rates? We have performed a number of controls (which are not presented here for reasons of space) and are convinced that none of these explanations is possible.

The fact that firing rates become more consistent over time is in keeping with our hypothesis that preparatory activity is optimized before movement generation. Indeed, given the results in figure 5, it seems plausible that the NV provides a rough signature of motor preparation. The NV declines following target onset, when motor preparation is thought to be occurring, and remains at a rough plateau during the delay, when motor preparation is thought to be on ‘hold’. The NV declines further following the go cue, suggesting that motor preparation is finalized before movement onset. Our interpretation would be that the height of the NV indicates the approximate degree of motor preparation yet to be accomplished. Shortly after target onset, firing rates are frequently far from their mean, and thus presumably far from optimal. If the go cue arrives then, it will take time to correct these errors and RT’s will be longer. By the time the NV has reached its

3

Computational and Neural Systems (COSYNE) Annual Meeting, Slide Presentation, 2005. In Press.Published abstract (pages 1-2) and supplementary material for COSYNE reviewers.

Neural variability and motor preparation Churchland et al.

plateau, firing rates are consistently near their mean (which we presume is near the optimal subspace). RT’s should therefore be shorter if the go cue arrives then. To test this idea, we repeated the experiment in one monkey (G) using three discrete delay periods (30, 130 and 230 ms), intended to interrupt motor preparation at varying degrees of completeness. The height of the NV at the time of the go cue was predictive of RT (figure 6). We further analyzed the data for the very short 30 ms delay period, and found that for trials with RT’s shorter than the median, the NV dropped more quickly (figure 7).

The above data argue that the NV provides a rough signature of the progress of motor preparation. However, it should be stressed that the measurement of variance is only a proxy to the true measure one would like to compute: the proximity of firing rates to the optimal subspace, on each trial, and as a function of time. We are not suggesting that the system is sensitive to firing rate variability itself. Quite the reverse, we are suggesting that the brain is capable of detecting when firing rates are not optimal, and delaying movement

onset until any errors can be corrected. We suspect that observing the NV will prove informative for a

variety of brain areas and tasks. The NV has an obvious drawback: it interprets consistency as accuracy, which is not guaranteed to be the case. However, the NV has a compensatory virtue: it allows one to estimate when firing rates become accurate (with the caveat mentioned above) even in the complete absence of any knowledge regarding the representational scheme used by that part of the brain. Many areas (e.g. infero-temporal and pre-frontal cortex) appear to use exceedingly complex representations, or perhaps cannot be sensibly said to employ any fundamental representation, making such a method particularly desirable. Experiment 3. The results of the above experiment suggest that the brain has a mechanism for monitoring the state of motor preparation, and delaying execution if further optimization is needed. To test this idea, we used intracortical microstimulation to disrupt PMd preparatory activity in two monkeys performing the delayed reach task. Microstimulation (333 Hz, 57 ms) was delivered at currents below the threshold for evoking movement, at a random time during the delay, or up to 50 ms following the go cue. We concentrate first on the effects omicrostimulation delivered just after the go cue (e.g. around the time that visual cortex begins to respond to the go cue). Microstimulation at this time had a very minimal impact on trajectory of the executed movement: both endpoint accuracy and peak velocity were very similar

to that for the unstimulated trials (data not shown). However, microstimulation just after the go cue considerably increased RT, as shown for one example site in figure 8. Across all sites in PMd (30 and 73 for monkey A and B), the mean increase in RT was 26 and 20 ms (p<10

f

.

-6 for each). In contrast, microstimulation delivered more than 100 ms earlier had a minimal impact on RT (increases of 7 and 2 ms), consistent with the idea that motor preparation is re-optimized in ~100 ms.

275

350

mea

n RT

(ms)

1 1 3NV at go cue

30 ms delay

230 ms

130 ms

Figure 6. Reaction time versus the normalized variance, at the time of the go cue, for three delay-period durations.

1.0

1.4

NV

100 ms

5

-5% d

iffer

ence

target

go

short

long RT

Figure 7. The NV for trials with a 30 ms delay period, computed separately for trials with RT’s shorter and longer than the median. The black trace at bottom shows the mean percent difference (with SEM, p<0.01 at the lowest point, p<0.05 collapsed across the 200 ms following the go cue). The blue trace shows the mean percent difference in firing rate.

Figure 8. Effect of microstimulation for one PMd site. Mean hand velocity (± SEM) as a function of time for trials with no stimulation (black), stimulation delivered just after the go cue (red) and stimulation delivered more than 100 ms before the go cue. (green). Data are combined across delays > 100ms.

go 300RT (ms)

0

100

hand

vel

ocity

(cm

/s)

stimstim In the absence of microstimulation, longer delay periods led to shorter RT’s (figure 9, black

4

Computational and Neural Systems (COSYNE) Annual Meeting, Slide Presentation, 2005. In Press.Published abstract (pages 1-2) and supplementary material for COSYNE reviewers.

Neural variability and motor preparation Churchland et al.

symbols), as has previously been reported (2, 5). However, the advantage of a longer delay was greatly reduced by microstimulation delivered just before the go cue. This finding is consistent with the hypothesis that microstimulation disrupts motor preparation, and therefore reduces the advantage of a long delay period.

0-100 100-400

270

300

0-100 100-400

210

240

Mea

n R

T (m

s)

Monkey A Monkey B

Delay (ms) Delay (ms)

The effect of microstimulation on RT was largely specific to PMd. Statistically significant effects were occasionally evoked from primary motor cortex (M1), but overall there was very little effect (12 and 32 sites, mean RT increases of 4 and 2 ms for post-go microstimulation, p>0.1). The effect of microstimulation was also largely specific to reaches; saccadic RT’s were increased only slightly (8 and 2 ms, p<0.001 and p>0.1). These findings are consistent with our interpretation that microstimulation delays RT by disrupting reach-specific preparatory activity, which is more prevalent in PMd than in M1. The finding that reaches are delayed, but accurate, supports the hypothesis that the brain is able to monitor preparatory activity, and delay movement when corrections are necessary.

It has previously been reported that transcranial shocks (both electrical and magnetic) of human motor cortex increase RT(17). The mechanism responsible appears to be massive cortical inhibition following the shock - as revealed by a period in which tonic EMG is suppressed(18). This cortical inhibition is thought to directly prevent movement generation(17, 19, 20). Consistent with this interpretation, the effects of transcranial shocks tend to be greatest over M1 (21). In contrast, we found very little effect of microstimulation in M1 Thus, despite their superficial similarity, the effect of transcranial stimulation and direct cortical microstimulation on RT appear to involve different mechanisms, with the former most likely inhibiting movement generation, and the latter most likely disrupting movement preparation. Conclusions. The above experiments provide converging evidence that motor preparation involves the optimization of preparatory activity in the motor cortices, particularly in PMd. The results of this optimization appear to be slightly different from trial to trial, at least partly explaining the copious variability inherent to voluntary movement. The process of optimization appears to consume time (~100-200 ms), possibly explaining the long RT of voluntary movements (as compared to cortical reflexes). Furthermore, optimization appears to take variable time, possibly explaining the considerable variability of natural RT’s. A corollary of this last finding is that the brain appears able to detect when motor preparation is inaccurate, and to delay movement until errors can be corrected, resulting in longer RT’s on some trials than others. References. 1. J. Tanji, E. V. Evarts, J Neurophysiol 39, 1062 (Sep, 1976). 2. D. A. Rosenbaum, J Exp Psychol Gen 109, 444 (Dec, 1980). 3. M. Weinrich, S. P. Wise, K. H. Mauritz, Brain 107 (Pt 2), 385

(Jun, 1984). 4. M. Godschalk, R. N. Lemon, H. G. Kuypers, J. van der Steen,

Behav Brain Res 18, 143 (Nov-Dec, 1985). 5. A. Riehle, J. Requin, J Neurophysiol 61, 534 (Mar, 1989). 6. D. J. Crammond, J. F. Kalaska, J Neurophysiol 84, 986 (Aug,

2000). 7. A. Riehle, J. Requin, Behav Brain Res 53, 35 (Feb 26, 1993). 8. M. Hirayama, M. Kawato, M. I. Jordan, J Mot Behav 25, 162

(Sep, 1993). 9. R. M. Kelly, P. L. Strick, Prog Brain Res 143, 449 (2004). 10. R. M. Kelly, P. L. Strick, J Neurosci 23, 8432 (Sep 10, 2003). 11. E. Todorov, M. I. Jordan, Nat Neurosci 5, 1226 (Nov, 2002). 12. R. J. van Beers, P. Haggard, D. M. Wolpert, J Neurophysiol 91,

1050 (Feb, 2004).

13. D. J. Tolhurst, J. A. Movshon, A. F. Dean, Vision Res 23, 775 (1983).

14. M. Gur, A. Beylin, D. M. Snodderly, J Neurosci 17, 2914 (Apr 15, 1997).

15. W. Bair, L. P. O'Keefe, Vis Neurosci 15, 779 (Jul-Aug, 1998). 16. B. B. Averbeck, D. Lee, J Neurosci 23, 7630 (Aug 20, 2003). 17. B. L. Day et al., Brain 112 (Pt 3), 649 (Jun, 1989). 18. S. A. Wilson, R. J. Lockwood, G. W. Thickbroom, F. L.

Mastaglia, J Neurol Sci 114, 216 (Feb, 1993). 19. U. Ziemann, F. Tergau, J. Netz, V. Homberg, Brain Res 744,

32 (Jan 2, 1997). 20. M. Hallett, Nature 406, 147 (Jul 13, 2000). 21. J. L. Taylor, D. S. Wagener, J. G. Colebatch,

Electroencephalogr Clin Neurophysiol 97, 341 (Dec, 1995).

Figure 9. Effect of microstimulation delivered in the 100 ms before the go cue. RT is plotted versus the delay-period duration for unstimulated (black) and stimulated (red) trials. Error bars show SEM. All main effects and interactions were statistically significant (ANOVA, p<0.05). Data are collapsed across all PMd sites that produced a >10 ms increase in RT for post-go stimulation.

5

Computational and Neural Systems (COSYNE) Annual Meeting, Slide Presentation, 2005. In Press.Published abstract (pages 1-2) and supplementary material for COSYNE reviewers.