Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Biol. Cybern. 88, 360–373 (2003) DOI 10.1007/s00422-002-0385-3 Ó Springer-Verlag 2003 Spike synchronization and firing rate in a population of motor cortical neurons in relation to movement direction and reaction time F. Grammont , A. Riehle Center for Research in Cognitive Neuroscience, CRNC – CNRS, Marseille, France Received: 16 January 2002 / Accepted in revised form: 26 November 2002 / Published online: 7 April 2003 Abstract. We studied the dynamics of precise spike synchronization and rate modulation in a population of neurons recorded in monkey motor cortex during performance of a delayed multidirectional pointing task and determined their relation to behavior. We showed that at the population level neurons coherently synchronized their activity at various moments during the trial in relation to relevant task events. The comparison of the time course of the modulation of synchronous activity with that of the firing rate of the same neurons revealed a considerable difference. Indeed, when synchronous activity was highest, at the end of the preparatory period, firing rate was low, and, conversely, when the firing rate was highest, at movement onset, synchronous activity was almost absent. There was a clear tendency for synchrony to precede firing rate, suggesting that the coherent activation of cell assemblies may trigger the increase in firing rate in large groups of neurons, although it appeared that there was no simple parallel shifting in time of these two activity measures. Interestingly, there was a systematic relationship between the amount of significant synchronous activity within the population of neurons and movement direction at the end of the preparatory period. Furthermore, about 400 ms later, at movement onset, the mean firing rate of the same population was also significantly tuned to movement direction, having roughly the same preferred direction as synchronous activity. Finally, reaction time measurements revealed a directional preference of the monkey with, once again, the same preferred direction as synchronous activity and firing rate. These results lead us to speculate that synchronous activity and firing rate are cooperative neuronal processes and that the directional matching of our three measures – firing rate, synchronicity, and reaction times – might be an effect of behaviorally induced network cooperativity acquired during learning. Correspondence to: A. Riehle (e-mail: ariehle@lnf.cnrs-mrs.fr, Tel.: +33-491-164329, Fax: +33-491-774969) Present address: Istituto di Fisiologia Umana, Università di Parma, Via Volturno 39, 43100 Parma, Italy 1 Introduction Reduction of uncertainty is one of the basic principles for understanding the mechanisms underlying preparation for action (Requin et al. 1991). In this context, (un)certainty is equivalent to information about the required motor response. Movement preparation can be studied best by manipulating information about various aspects of the movement. For that purpose two signals are presented successively during each trial and separated by an instructed delay, the preparatory period. The first, or preparatory, signal provides prior information about (aspects of) the movement, which must be executed after the occurrence of the second, or response, signal. Two main categories of information may be manipulated. On the one hand, providing prior information about spatial and/or kinematic parameters of the movement, e.g., direction, extent, force, etc., reduces uncertainty such that it leads to a significant reduction in reaction time (Rosenbaum 1980; Bonnet et al. 1982; Lépine et al. 1989; Riehle and Requin 1989, 1995; Riehle et al. 1994). On the other hand, manipulating temporal aspects of the task by systematically varying the duration of the preparatory period has been shown to efficiently alter the preparatory state of the subject (Bertelson and Boons 1960; Bertelson 1967). When presenting a finite number of durations of the preparatory period at random, but with equal probability, reaction time decreases with increasing duration (for a review see Requin et al. 1991). Indeed, as time goes on during the trial and the response signal is not presented at the first possible moment, the probability for its occurrence increases with each subsequent possible moment (Riehle et al. 1997). It is commonly accepted that behavioral and cognitive processes are reflected in the (mean) discharge rate of neurons (Barlow 1972; Shadlen and Newsome 1994). In the framework of movement preparation, it has been shown that motor cortical neurons selectively change their activity long before movement execution in relation to prior information about various parameters such as direction (Weinrich and Wise 1982; Weinrich et al. 1984; Riehle and Requin 1989; Bastian et al. 1998), 361 force (Riehle et al. 1994; Riehle and Requin 1995), or movement extent (Riehle and Requin 1989; Kurata 1993; Riehle et al. 1994). In parallel, another concept emerged claiming that computational processes in neocortical areas could also rely on the relative timing correlation of spiking discharges among neurons within functional groups (Abeles 1982, 1991; Gerstein et al. 1989; Singer 1999), commonly called cell assemblies (Hebb 1949; for a review see Fujii et al. 1996). Experimental evidence for such precise temporal relations among spiking activities of multiple neurons was provided essentially in sensory areas (visual: reviewed in Singer and Gray 1995; auditory: Ahissar et al. 1992; DeCharms and Merzenich 1996; somatosensory: Nicolelis et al. 1995; frontal: Aertsen et al. 1991; Vaadia et al. 1995). Recently it has been shown that in motor cortical areas spike synchronization can also occur in relation to signal expectancy (Riehle et al. 1997, 2000). Furthermore, spike synchronization was observed during sensorimotor transformation between neurons that were classified, on the basis of their changes in firing rate, as being functionally involved in different processes, e.g., preparation- or execution-related (Grammont and Riehle 1999). Finally, it has been shown that spike synchronization of motor cortical neurons, along with their discharge rate, carried about 10% more information about movement direction than discharge rate alone (Hatsopoulos et al. 1998). Studies on neuronal coding has stimulated an important discussion in recent years among experimentalists (Abeles et al. 1993; Shadlen and Newsome 1994; Softky 1995; Singer and Gray 1995; Roskies 1999) and theoreticians (Von der Malsburg 1981, 1995; Abeles 1982, 1991; Aertsen et al. 1994, 1995; Diesmann et al. 1999). Of particular interest here are debates currently in progress for determining the implication of a temporal code in cognitive processes (for reviews see Fujii et al. 1996; Roskies 1999). As two modes of neural coding, i.e., rate code and temporal code, may in fact be considered to be related to cognitive processes at various levels, our integrative approach tends to explore their complementary involvement in one of the same processes – here preparation for action. To that end we simultaneously recorded the activity of up to seven single neurons in the primary motor cortex of a monkey in the context of a delayed multidirectional pointing task. Both prior information about movement direction and the duration of the instructed delay were systematically manipulated. We used the modified version of the unitary event analysis for detecting synchronized spiking activities in pairs of neurons (Grün 1996; Grün et al. 1999, 2002a,b). Basically this technique allows one to determine epochs containing spike coincidences that violate the assumption of independence of the participating neurons. The interdependence is then interpreted as a signature of a functional cell assembly (Aertsen et al. 1991). The statistical null-hypothesis is formulated on the basis of the individual firing probabilities and allows one to calculate the number of expected coincidences. The statistical significance of the measured number of coincidences is evaluated by comparing it with the expected number. By using this analysis in sliding windows, we can deal with nonstationarities of firing rates and determine epochs of significant synchronized activity along the trials. The temporal precision of spike synchronization is obtained by additionally varying the allowed coincidence width in the analysis between 1 to 20 ms (Grün et al. 1999). If cell assemblies are involved in cortical information processing, they should be activated in systematic relation to the behavioral task (Riehle et al. 1997, 2000; Grammont and Riehle 1999). Thus, in order to clarify the contribution of cell assemblies to cognitive motor processes, we quantified data from many pairs of neurons by calculating the probability of significant synchronization (Grammont 2001). This provides a measure describing the evolution both in time and temporal precision of synchronous spiking activity at the level of an entire population of single neurons. We then compared the dynamics of synchronized activity of this population of neurons with its mean firing rate, according to movement direction and reaction time. Preliminary results have been presented in abstract form (Grammont and Riehle 2000). 2 Material and methods 2.1 Behavioral procedure A male Rhesus monkey (Macaca mulatta) was trained to perform a delayed multidirectional pointing task (Riehle et al. 2000). It was cared for in the manner described in the Guiding Principles in the Care and Use of Animals of the American Physiological Society and in French government regulations. The animal sat in a primate chair in front of a vertical panel on which seven touchsensitive, light emitting diodes (LEDs) were mounted, one in the center and six placed equidistantly on a circle around it (Fig. 2, inset). The monkey had to initiate a trial by touching with the left hand the central target when it was lit in yellow (start trial, ST, Fig. 1a). After a fixed delay of 500 ms during which the monkey had to continue pressing the target, two signals were presented successively, separated by a variable time interval. The first, the preparatory signal (PS), consisted of the illumination of one of the peripheral targets in green. It indicated the target for the upcoming movement. After a delay of either 600 or 1200 ms, presented at random with equal probability, during which the animal had to continue to press the central target, the illuminated peripheral target turned red, serving as response signal (RS) and target to be pointed. During the first 600 ms of the preparatory period of each trial, the probability for the response signal to occur at 600 ms was 0.5 (RS1). Once this moment passed without signal occurrence, conditional probability changed to 1 (RS2). The monkey was rewarded by a drop of juice at the end of each correct trial. It completed 300 trials during one session including 12 trial types, 6 movement directions in combination with 2 durations of the preparatory period. Reaction time was defined as the delay between the 362 electrodes, outer diameter 80 lm, impedance: 2–5 MX at 1000 Hz). The electrodes were arranged in a circle, one electrode in the middle and six around it, equally spaced 330 lm apart. From each electrode, electrical signals were amplified and band-pass filtered from 300 Hz to 10 kHz. Using a window discriminator, spikes from only one single neuron per electrode were then isolated. Neuronal data along with behavioral events (e.g., occurrences of signals and performance of the animal) were stored on a PC for offline analysis with a time resolution of 1 kHz. Electromyographic (EMG) activity was recorded from nine selected muscles of the active arm and shoulder as well as the trunk during representative sessions by using surface electrodes on the skin. Prior to recordings hairs on the skin were carefully removed to provide electrical contact. The muscles were the deltoidius, trapezius, triceps, biceps, carpi radialis, carpi ulnaris, policis longus, bracchio radialis, and pectoralis. A –500 ST B 0 600 PS RS1 or ES 1200 ms RS2 reaction time (ms) 150 140 130 120 110 4 5 6 1 directions 2 C 3 D 5 6 5 6 2.3 Data analysis 4 1 3 2 4 1 3 2 Fig. 1. Experimental design of the task. a Schematic representation of a trial. For details see Sect. 2.1. ST: start trial; PS: preparatory signal; ES: expected (response) signal; RS: response signal, RS1 after a short trial and RS2 after a long one. b Mean reaction times (ms standard errors, solid curve) as a function of movement direction in long delay trials. The cosine fit (dash-dotted) revealed a systematic, statistically significant tuning to movement direction (r ¼ 0:79, p < 0:05). The behavioral preferred direction (PD), corresponding to the shortest reaction time determined by the cosine fit, was between target 6 and 7 (PD ¼ 6:7). Directional modulation index I ¼ 0:11. c and d Distributions of preferred directions of the significantly tuned neurons in the selected data set (24/33 neurons, c) and the entire population (301/ 396 neurons, d) occurrence of the response signal and the release of the central target. 2.2 Surgical procedures and recording technique After training, the animal was prepared for multiple single-neuron recordings. A cylindrical stainless steel recording chamber (inner diameter: 15 mm) was implanted over the contralateral (right) primary motor cortex under aseptic conditions and general halothane anesthesia (< 2:5% in air). A stainless steel T-bar was cemented to the skull to fixate the animal’s head during recording sessions. Before and after surgery antibiotics and analgesics were administered. In order to record extracellular single-neuron activity, a multielectrode microdrive (Reitböck system, Thomas Recording, Giessen, Germany; Mountcastle et al. 1991) was used to transdurally insert independently from each other seven microelectrodes (quartz-insulated platinum-tungsten 2.3.1 Unitary event analysis. Dynamic changes in the temporal relations between the occurrences of spikes in sets of simultaneously recorded pairs of neurons were analyzed offline on a UNIX workstation with Matlab (The MathWorks, Inc.) by using an extension of the unitary event method (Grün 1996), the multiple shift method (Grün et al. 1999). It treats the data in their (original) high time resolution (1 ms). Such a procedure allows one to search for coincidences with various coincidence widths. Technically, coincident spikes are detected in pairs of neurons by shifting the spike trains against each other over the range of allowed coincidence widths (ranging from 1 to 20 ms) and integrating the number of exact coincidences (on the time resolution of the data) over all shifts. In order to evaluate the significance of the detected coincidences (measured coincidences), the outcomes are compared to the expectation. Under the null-hypothesis that neurons fire independently from each other, the expected number of occurrences (expected coincidences) and its probability distribution can be estimated on the basis of the single-neuron firing probability. Expectation is calculated as the product of the marginal probability of firing and then summed for all possible shifts. The statistical significance of a positive or negative difference between measured and expected coincidences can be assessed from a Poisson distribution (with the mean set to the expected coincidence value) as the cumulative probability P of observing a larger or smaller number of coincidences than expected by chance (Fig. 2 in Grün et al. 2002a). The larger the number of excessive coincidences, the closer P is to 0. Conversely, the larger the number of lacking coincidences, the closer its complement, 1 P , is to 0, while P approaches 1. Those occurrences of coincident spikes that exceed the significance level of 5% were called unitary events. For a better visualization (Fig. 3c), we use a logarithmic 363 events at each instant in time along the trials over the various pairs of neurons and movement directions (Riehle et al. 2000; Grammont 2001). Technically, for each pair of neurons we constructed each movement direction and each coincidence width (from 1 to 20 ms) a binary vector that indicates by ‘‘1’’ a significant sliding window of unitary events and by ‘‘0’’ a nonsignificant one. In Fig. 3d, an example of such a binary vector is provided for one pair of neurons during movement direction 6 by selecting a coincidence width of 2 ms. By averaging all binary vectors determined for a given coincidence width and behavioral condition, we calculated the probability of significant synchronization as a function of time and coincidence width. For brevity, in what follows we call this measure synchronicity. 2.3.3 Directional selectivity. In order to test directional selectivity, mean data of synchronicity, firing rate, and behavioral reaction times were fitted to a model by using an adapted method originally proposed by Georgopoulos et al. (1982). A first-degree periodic (sinusoidal) regression to be fitted is Fig. 2. Electromyographic (EMG) activity was recorded from nine muscles of arm, shoulder, and trunk of the active side during representative sessions with surface electrodes on the skin. The muscles were the 1: deltoidius, 2: trapezius, 3: triceps, 4: biceps, 5: carpi radialis, 6: carpi ulnaris, 7: policis longus, 8: bracchio radialis, and 9: pectoralis. PS: preparatory signal, ES: expected (response) signal, RS: (actual) response signal, Mvt: movement onset. The inset shows the spatial configuration and numbering of the targets (movement directions). Movements were to be started from the center circle function of the two: log10 ½ð1 P ÞP (surprise measure, Palm et al. 1988). To deal with nonstationarities through time in the firing rate of the neurons, temporal relations among spikes were analyzed in time segments by using a sliding boxcar window of 100 ms, which was shifted in steps of 5 ms along the data. This sliding window procedure was applied to each trial and the data of corresponding segments in all trials were then analyzed as one stationary data set. Only neurons were selected for analysis that reached the following criteria: (1) a lowest firing rate of more than seven impulses/second (Roy et al. 2000), (2) stationarity of rate changes across trials (Grün et al. 2002b; this issue), (3) a minimum of about 20 trials per condition, and (4) of course at least two neurons had to fulfill these criteria in the same recording session in order to constitute a pair. All these criteria were applied to each movement direction. As a consequence, for one single pair of neurons not all movement directions were necessarily selected for further analysis so that the sample size is smaller than the number of selected pairs times the number of movement directions. 2.3.2 Quantification of synchronized spiking activity. In order to quantify synchronous activity at the population level, we calculated the probability of observing unitary y ¼ b0 þ b1 sin h þ b2 cos h ð1Þ where b0 , b1 and b2 are regression coefficients. The least square unbiased estimators of these coefficients for six movement directions are b0 ¼ ðy1 þ y2 þ y3 þ y4 þ y5 þ y6 Þ=6 ð2Þ where y1 . . . y6 are the mean data for each of the six movement directions. b1 ¼ 1 pffiffiffi 3ðy2 y3 þ y5 þ y6 Þ 2 1 1 1 1 1 y1 þ y2 y3 y4 y5 þ y6 b2 ¼ 3 2 2 2 2 ð3Þ ð4Þ Statistical significance of the tuning is determined by calculating a linear correlation coefficient r between the empirical data and the cosine fit. The preferred direction, PD, is calculated using the circular mean of the fitted sine wave and expressed in fractions of target numbers. Finally, a directional modulation index, I, is determined by calculating a proportional increase of the studied data sample at the preferred direction over its overall mean b0 (Georgopoulos et al. 1982) b0 I ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi b21 þ b22 ð5Þ 3 Results 3.1 Behavioral results and EMG data Mean reaction times were calculated for each of the 12 trial types on the basis of correct trials in all recording sessions. The sample size (n ¼ 118, see below) for each 364 coinc/sec B spikes/sec Spike Rates, neuron 1 (solid), neuron 4 (dashed) A 50 0 Coincidence Rates: measured (solid), expected (dashed) 20 10 C J–surprise 0 Significance Level 2 0 –2 Binary Vector (zero one) D sig (1) n–sig (0) PS 300 ES 900 RS yo049–14– 6 Fig. 3. Dynamic changes in synchronous spiking activity of a pair of neurons during the execution of one type of trials (long preparatory period, direction 6). For calculation, a sliding boxcar window of 100 ms was shifted along the spike trains in 5-ms steps. The allowed coincidence width was 2 ms. Since data were aligned trial by trial to movement onset (vertical line in all plots), the occurences of the signals (gray bars) were plotted with respect to the mean reaction time, the width of the bars corresponding to its standard deviation (RT ¼ 89 47 ms). PS: preparatory signal, ES: expected (response) signal, RS: response signal. Time is running along the x-axis and indicated in milliseconds. a Firing rate of the two neurons in spikes/ second. b Measured (solid) and expected (dashed) coincidence rates are shown in coincidences/second. c For each sliding window the statistical significance (joint-surprise value) was calculated for the difference between measured and expected coincidence rates. The result of each window was placed in its center. Whenever the significance value exceeded the threshold (upper dashed line, p ¼ 0:05), this defined an epoch in which significantly more coincidences occurred than expected by chance. Occasionally, this value dropped below the lower dashed line, thus indicating epochs in which significantly (p < 0:05) fewer coincidences occurred than expected by chance. d A binary vector is plotted on which ‘‘1’’ indicates a significant sliding window of unitary events and ‘‘0’’ a nonsignificant one. The ‘‘1’’ and ‘‘0’’ entries were plotted in the center of each corresponding sliding window of 100 ms. The binary vectors were subsequently used for the quantification analysis presented in Fig. 4. Synchronous activity was at chance level during the first 600 ms of the preparatory period and then started to increase. It modulated for the remaining period with increasing amplitude to reach a maximum just before the presentation of the response signal. During movement onset neuronal activity was considered to be antisynchronized, i.e., there were significantly fewer coincidences than expected delay duration corresponds to the number of selected pairs times the number of movement directions after removal of individual data sets that did not reach the criteria defined in Sect. 2.3.1. Since two different durations of the preparatory period were presented at random with equal probability, the response signal could occur either 600 or 1200 ms after the preparatory signal. Mean reaction times in relation to the duration of the preparatory period are shown in Table 1. Reaction times were significantly shorter when they were produced after a long preparatory period [RT(2)], i.e., when the probability for the response signal to occur was 1, than after a short one [RT(1)], i.e., when the probability of signal occurrence was 0.5. The difference of 130 ms between short and long trial reaction times over all movement directions is highly statistically significant (t-test of Student: t ¼ 31:59, df ¼ 234, p 0:0001). The same is true for the difference in reaction time between right-sided and left-sided movements after long delays (t ¼ 2:82, df ¼ 116, p < 0:01) and short delays (t ¼ 7, df ¼ 116, p << 0:001). Individual reaction times during each movement direction and delay duration are indicated in Table 1. In Fig. 1c, a cosine function was fitted (dash-dotted line) to the reaction time data for each movement direction in long delay trials, revealing statistical significance of the fit (r ¼ 0:79, p < 0:05). Furthermore, the cosine fit allows one to determine the behavioral preferred direction, i.e., the movement direction in which reaction time was shortest (PD ¼ 6:7). However, although reaction times were significantly tuned to movement direction, the directional selectivity index was rather low (I ¼ 0:11). We recorded electromyographic (EMG) activity from task-related muscles of the active arm and shoulder. Figure 2 shows a typical example of the activity of nine muscles during long trial performance in each movement 365 Table 1. Mean reaction times (RT ) and standard deviations (SD), in milliseconds, were calculated for each movement direction and each duration of the preparatory period (PP ) from the means in each recording session per neuron pair Direction Short PP RT(1) Long PP RT(2) All directions Right movements (directions 6-1-2) Left movements (directions 3-4-5) Direction 1 Direction 2 Direction 3 Direction 4 Direction 5 Direction 6 254 ± 29 238 ± 19 124 ± 36 115 ± 27 269 ± 29 132 ± 41 241 238 271 284 252 236 111 113 145 131 121 121 ± ± ± ± ± ± 22 18 26 29 24 18 ± ± ± ± ± ± 23 23 41 44 36 35 direction. It can clearly be seen that during the preparatory period, and particularly at the moment when a response signal was expected (ES), no overt change in activity could be detected in any muscle. 3.2 Neuronal results 3.2.1 Data set. Among 396 neurons recorded in the primary motor cortex, only 33 were selected from 14 recording sessions by severely applying the criteria described in Sect. 2.3.1 for constituting 22 pairs. A single neuron could participate in several pairs. The main reason for the small number of selected neurons is that most of the neurons phasically changed their activity in relation to signal occurrence and/or signal expectancy, while returning to almost zero spikes/second between bursts, thereby making the unitary event analysis impossible (see Sect. 2.3.1). Furthermore, all selected neurons changed their activity in close temporal relationship to movement performance. In order to preserve stationarity across trials, spiking activities were therefore aligned, trial-by-trial, to movement onset (Fig. 8 in Grün et al. 2002b). To qualify the discharge patterns of the neurons, their spiking activities were tested for directional selectivity by fitting a cosine to the mean firing rates determined during movement execution, defined as the period between the occurrence of the response signal and touching the requested target. In 24 neurons (73%), directional selectivity was statistically significant (p < 0:05). The distribution of preferred directions of these neurons is shown in Fig 1c. The difference, however, between preferred directions of directionally selective neurons within a pair extended over the whole possible range of movement directions from 6 to 168 . By comparison, directional selectivity was statistically significant during movement execution in 76% of the neurons of the entire population (301/ 396). The distribution of the preferred directions of the significantly tuned neurons of the whole population is represented in Fig. 1d. Both distributions proved to be homogeneously distributed over the whole range of movement directions by calculating the circular variance (selected sample: S24 ¼ 0:93, whole population: S301 ¼ 0:97, Mardia 1972). For each pair of neurons and in each behavioral condition, significant synchronous spiking activities were calculated individually for each coincidence width between 1 and 20 ms, as described in Sect. 2.3.2. Only data obtained during the long preparatory period were taken into account for the analysis at the population level, because there is no a priori difference between the short preparatory period and the first half of the long one; during this part of the trial the animal did not know whether or not a response signal would occur at 600 ms. Figure 3 shows a typical example of the distinct modulation of synchronous activity during the trial. 3.2.2 Modulation of synchronicity as a function of time and coincidence width at the population level. In order to test whether statistically significant synchronous activity during particular epochs of the trial was reproducible over the entire population of neurons, we quantified individual data (see Sect. 2.3.2). The probability for obtaining significant synchronization calculated among all pairs of neurons and all movement directions, thus obtaining a sample size of n=118, is shown in Fig. 4a. Probability was calculated as a function of both time and coincidence width. If neuronal synchronization plays a role in information processing, it should modulate systematically through time and coincidence widths rather than being homogeneous. Indeed, distinct increases in synchronicity could be observed at three specific epochs of the task: (1) around the preparatory signal, (2) precisely after the moment when the response signal was first expected but did not occur, and (3) about 300 ms before the actual occurrence of the response signal. By comparing the first with the second half of the preparatory period, synchronicity appeared to be much higher during the second than the first half, i.e., when the probability for the response signal to occur passed from 0.5 to 1. However, synchronicity dropped down to almost zero during movement execution. Remember that the higher synchronicity, i.e., the higher the probability of being significantly synchronized, the more pairs of neurons synchronized significantly their activity. Note that the precision of synchronous activity, i.e., the range of coincidence widths, also varied over time. 3.2.3 Modulation of firing rate at the population level. The mean firing rate was calculated for the same population of neurons and averaged over all movement directions (Fig. 4b). There was a slight tonic increase in activity during the first part of the preparatory period followed by two successive phasic modulations during its second part. Indeed, activity peaked first about 150 ms after the first response signal was expected but did not occur, reaching a first maximum of about 27 spikes/second. Firing rate then peaked a second time at movement onset and reached its absolute maximum of 44 spikes/second, showing the typical motor activity for primary motor cortical neurons. Interestingly, when synchronicity was highest (Fig. 4a), about 300 ms before the occurrence of the 366 coincidence width (ms) A Probability of Significant Synchronization 20 15 10 5 0 0 300 0.04 spikes / s B 0.06 600 900 Probability (max=0.19) 0.08 0.1 0.12 1200 0.14 MVT 0.16 0.18 Mean Firing Rate 45 35 25 15 PS 300 ES 900 response signal, firing rate was low (Fig. 4b), and, conversely, when the discharge was highest at movement onset, synchronicity was lowest. In fact, the increase of synchronicity tended to precede the increase in firing rate, although not in a systematic temporal relationship. 3.2.4 Comparison between synchronicity and mean firing rate as a function of movement direction. In Fig. 5, the same data as in Fig. 4a are presented separately for each movement direction. It is clear that the modulation of the probability of significant synchronization was not homogeneously distributed over movement directions. For instance, during the second part of the preparatory period the probability of significant synchronization was highest for direction 6 but very low for the opposite direction 3. Synchronicity varied between these two extremes in an apparent systematic manner. In fact, a further analysis realized during the second part of the preparatory period demonstrates that these variabilities observed in terms of both synchronicity and firing rate RS MVT Fig. 4. Quantification of correlated activity of all pairs of selected neurons pooled over all six movement directions (n ¼ 118). a For each pair of neurons and each movement direction a matrix was calculated containing entries of 0s and 1s (binary vectors, Fig. 3d) based on the statistical significance for spike synchronization (parameters: sliding window of 100 ms, shift offset 5 ms, coincidence widths ranging from 1 to 20 ms). All obtained matrices were then pooled to calculate the probability matrix. It contains per entry position the probability for being significantly synchronized across behavioral conditions (from dark blue to dark red probabilities ranging from 0.02 to 0.19, see color bar). b The averaged mean firing rate from the same neuronal population (also calculated within sliding windows of 100 ms width, which were shifted in 5-ms steps along the trial). PS: preparatory signal, ES: expected (response) signal, RS: (actual) response signal, MVT: movement onset. Since all data are aligned, trial by trial, to movement onset, signal occurrences varied from trial to trial as a function of reaction time. The width of the gray bars indicating the signals corresponds to the standard deviation of reaction times followed a systematic law (Fig. 6). We calculated the tuning curves of the two types of neuronal activity at two distinct moments, corresponding, first, to the peak of mean synchronicity (dotted line at 300 ms before the response signal) and, second, to the peak of the mean firing rate (dash-dotted line shortly after movement onset). Both the probabilities of significant synchronization and the mean firing rates were tuned, but at two different moments. Indeed, synchronicity was significantly tuned during the preparatory period (Fig. 6c; cosine fit: r ¼ 0:92, p < 0:01), whereas firing rate showed no tuning at the same moment (Fig. 6e). In contrast, firing rate was significantly tuned in relation to movement onset (Fig. 6f; cosine fit: r ¼ 0:93, p < 0:01), whereas synchronicity was very low and did not modulate with movement direction at this time (Fig. 6d). In both cases, when synchronicity and firing rate were significantly tuned, the preferred direction (PD) was roughly the same, i.e., less than a target distance apart, that is, 6.8 for synchronicity long before 367 Direction 5 (max=0.3) Direction 6 (max=0.37) 20 20 15 15 10 10 5 5 0 0 coincidence width (ms) Direction 4 (max=0.23) Direction 1 (max=0.32) 20 20 15 15 10 10 5 5 0 0 Direction 3 (max=0.24) Direction 2 (max=0.3) 20 20 15 15 10 10 5 5 0 PS 0 ES 0.1 RS MVT 0.2 0.3 0 0 300 0 600 0.1 900 0.2 1200ms 0.3 Probability of Significant Synchronization Fig. 5. Probability of significant synchronization for each movement direction during long delay trials. The same type of analysis as in Fig. 4a was applied to the same population data for each movement direction separately. The location of each matrix in the figure corresponds to the location of the respective targets (Fig. 2, inset). Note, however, that the probability for being significantly synchronized was much higher (color bar ranging from 0 to 0.3 is valuable for all six matrices) in individual data, reaching a maximal probability of 0.37 (direction 6). For the analysis, all data were aligned, trial by trial, to movement onset (MVT, black vertical lines). The red vertical lines correspond to the preparatory signal (PS), the expected (response) signal (ES), and the actual response signal (RS). Since for each movement direction mean reaction times were different (Table 1), signal occurrences are indicated accordingly movement onset (Fig. 6c) and 6 for the firing rate during movement onset (Fig. 6f). As an individual example and for compatibility with Fig. 7, the synchronicity tuning was also calculated for a single coincidence width of 6 ms. This analysis revealed virtually the same results as above, namely, a significant tuning during the preparatory period (PD ¼ 6:9, r ¼ 0:95, p < 0:001, I ¼ 1:1) and no tuning during movement onset (r ¼ 0:67). By fitting any function to the data, only the mean values determined for each movement direction are tested for their statistical significance in relation to the 368 Probability of Significant Synchronization A probability 0.3 0.2 0.1 0 Mean Firing Rate per Movement Direction B 50 Dir 6 Dir 1 Dir 2 Dir 3 Dir 4 Dir 5 spikes / s 40 30 20 10 ES –450 –300 During Preparatory Period C 0.1 MVT 0.3 probability 0.2 RS During Movement Onset D 0.3 probability –150 r=0.49 non–sel 0.2 0.1 PD=6.8 r=0.93 p<0.01) I=0.69 0 E 4 5 6 1 2 F 30 r=0.65 non–sel 25 20 5 6 1 directions 2 3 5 6 1 2 3 45 40 4 4 50 spikes / s spikes / s 0 3 PD=6 r=0.93 4 chosen model as, for instance, a cosine function. However, this does not indicate whether or not data samples, which are at the origin of the means, are significantly different from each other. In the next analysis we therefore tested if there was a statistically significant difference between neuronal activities during right-sided and left-sided movements (Fig. 7). In order to apply a v2 -test to the individual data from all neuron pairs, we have chosen for testing the difference in synchronicity a typical coincidence width of 6 ms (Fig. 4a and the individual calculation performed for the data presented in Fig. 6). In each sliding window of 100 ms, the proportions of ‘‘1’’ entries with respect to the sample size, corresponding to the number of significant data sets at 5 p<0.01 I=0.11 6 1 directions 2 3 Fig. 6. Comparison between synchronicity and firing rate in the same population of neurons as in Figs. 4 and 5 during the second part of long delay trials. a Probability of significant synchronization, averaged over all coincidence widths, for each movement direction separately. b Mean firing rate for the same data as in a. ES: expected (response) signal, RS: (actual) response signal, MVT: movement onset. Since all data were aligned, trial by trial, to movement onset, signal occurrences varied from trial to trial as a function of reaction time. The width of the gray bars indicating the signals corresponds to the standard deviation of reaction times. c–f Tuning curves (solid) as a function of movement direction and the corresponding cosine fits (dash-dotted) are shown. The tuning was calculated from the mean values of the probability of significant synchronization (c and d) and the corresponding firing rate (e and f) at two distinct moments during the trial indicated by the two vertical lines in a and b. Tuning curves at the moment indicated by the dashed lines are shown in c and e, whereas those indicated by the dash-dotted lines are shown in d and f. In each subfigure, the correlation coefficient r of the linear regression between the tuning curve and the cosine fit and its statistical significance is indicated. Furthermore, in subfigures in which the tuning is statistically significant (c and f), the preferred direction (PD) is indicated in fractions of target numbers as well as the directional modulation index I this epoch (Sect. 2.3.2), were compared for right-sided and left-sided movements, and their statistical significances were calculated. Only the data inside the marked boxes were significantly different (p < 0:05), two adjacent windows at about the moment when a response signal was expected, but did not occur, and ten adjacent windows at about 300 ms before its actual occurrence. In other words, within the entire population significantly more neurons synchronized significantly their activity during preparation for right-sided movements than during preparation for left-sided movements (Fig. 7a). In Fig. 7b, the mean firing rates were compared for the same right-sided and left-sided movements by applying, sliding window by sliding window, a t-test of 369 Probability of Significant Synchronization A Right movements Left movments probability 0.25 0.2 0.15 0.1 0.05 0 300 600 900 1200ms Mean Firing Rate B spikes / s 40 30 20 PS 300 ES 900 RS MVT Fig. 7. Comparison between synchronicity and firing rate in the same population of neurons as in Figs. 4 to 6 during the long delay trials in relation to right-sided (solid, n ¼ 58) and left-sided (dash-dotted, n ¼ 60) movements. a Probability of significant synchronization at a coincidence width of 6 ms. The difference in synchronicity between right-sided and left-sided movements was statistically significant inside the two boxes, i.e., when the response signal was expected and about 300 ms before its actual occurrence (chi square-test, v2 ranging between 5.05 and 11.25, df ¼ 1, p 0:05). b Mean firing rate in the same conditions as in a. The application of a t-test of Student for comparing individual data during right-sided and left-sided movements in each sliding window did not reveal any statistical significance Student to the samples of mean activities of all individual data sets. Although the mean firing rate was higher during preparation and execution of right-sided movements than left-sided movements, no statistical significance could be detected so far. 4 Discussion 4.1 Time course of synchrony and firing rate In this paper, we compare the time course of the modulation of precise spike synchronization with that of the mean firing rate in a population of motor cortical neurons in the behaving monkey during the performance of delayed multidirectional pointing movements. In general, significant synchronous activity reveals, at the population level, a clear structure (Fig. 4a). This suggests that motor cortical neurons systematically synchronized their activity in relation to relevant task events. Interestingly, the temporal pattern of synchronous activity is by no means predictable by inspecting the mean firing rate of the same neuronal population. The task was designed such that the animal could anticipate signal occurrences, including preparatory and response signals. As a matter of fact, a first peak in synchronicity preceded the occurrence of the preparatory signal, although with low temporal precision. A second one followed precisely at the moment when the first possible response signal was expected, but did not occur, preceding the first peak in firing rate. Finally, synchronicity reached a maximum at the end of the instructed delay, about 300 ms before the occurrence of the actual response signal, whereas the mean firing rate did so much later, at movement onset, an epoch where synchronous activity was almost absent. Although the task was temporally organized in periods of 600 ms, the dynamics of synchronicity did not follow a rhythmic temporal structure. Indeed, an increase in synchronous activity sometimes preceded a task event, sometimes followed it, depending on the underlying cognitive processes. Furthermore, the temporal precision of synchronous activity varied systematically during the task, from very low precision around the preparatory signal to highest temporal precision (2 to 3 ms) at the end of the preparatory period. Note that in individual pairs of neurons the precision of synchrony may reach values of 1 ms (Riehle et al. 2000; 2 ms in Fig. 3). Such a timing suggests that synchronous neuronal activity in motor cortex may be preferentially involved in early preparatory and motor cognitive processes, including signal expectancy (cf Riehle et al 1997), whereas the modulation in firing rate may control rather movement initiation and execution. In particular, the fact that during movement onset firing rate was highest and synchrony almost absent indicates that processes underlying movement initiation relate almost exclusively to changes in firing rate. An additional argument in favor of this hypothesis is the fact that EMG activity (Fig. 2) did not reveal any modulation during the preparatory period, indicating that neuronal activity was not involved in executive processes at this moment, i.e., cortical preparatory processes did not reach periphery. Furthermore, by comparing the time courses of synchronicity and mean firing rate in an entire population of neurons, it appeared that there was no simple parallel shifting in time of these two measures, as already described for individual pairs of neurons (Riehle et al. 2000). This makes it unlikely that the two coding schemes are tightly coupled by any kind of stereotyped transformation; they seem to obey rather different dynamics. There is, however, a clear tendency for synchrony to precede firing rate, suggesting that the coherent activation of cell assemblies may trigger the increase in firing rate in large groups of neurons. 4.2 Movement initiation In contrast to the data described by Hatsopoulos et al. (1998), we detected virtually no significant synchronization during movement onset, although firing rate was highest. This is true for both long (Fig. 4a) and short trials (not shown). The increase in firing rate in relation to movement onset suggests that the type of neurons in both studies was similar and rather movement related. In our data, however, peak synchrony occurred in long trials long before movement onset, about 300 ms before presentation of the response signal. This difference may be due to the experimental design. Hatsopoulos and colleagues (1998) employed a random 1-s to 1.5-s 370 instructed delay, whereas in our task the temporal constraints were precisely defined by presenting only two fixed durations at random. This means that the monkey could estimate the time of signal occurrences, although with different probabilities. During the second part of the preparatory period, the probability for the response signal to occur at a precise moment was 1. The first peak in synchronicity appeared right after the moment when the first response signal was expected but did not occur. This increase in the number of neurons synchronizing their activity is thus related to a purely internal – cognitive – event, the moment when the animal realized that the expected signal did not occur. The second and highest peak of synchronicity, which was also internally triggered, occurred about 300 ms before the actual response signal. In both tasks, directional information could already be processed during the waiting period, inciting the animal to prepare the movement in advance. However, movement initiation could be anticipated precisely in our task but not in that of Hatsopoulos et al. (1998). The fact that reaction time was significantly longer after a short delay – a situation in which the probability for the occurrence of the response signal and, thus, the probability to execute the requested movement was much lower (p ¼ 0:5) than after a long delay (p ¼ 1) – suggests that the monkey indeed prepared the movement in advance. In the context of the preparation paradigm, the preparatory coherent activation of cell assemblies, by way of synchrony, might trigger the increase in firing rate in large neuronal networks, which in turn communicate with the periphery for initiating the movement. 4.3 Movement direction We first analyzed the discharge patterns of individual neurons during each movement direction. The percentage of significant directional selectivity, defined by fitting a cosine function (73% of the selected data sample and 76% of the entire data set), corresponds to data described in the literature (e.g., Georgopoulos et al. 1982; Hatsopoulos et al. 1998). Although in the majority of pairs (14/22, 64%) the activities of both neurons were significantly tuned during movement execution, there was no systematic relationship between preferred directions of the two neurons forming the pair. Such an example was already shown by Grammont and Riehle (1999) for three simultaneously recorded neurons in a similar task during which the duration of the preparatory period was kept constant. Even though the firing rates of all three neurons were differently tuned to movement direction (see inset in Fig. 2 of Grammont and Riehle 1999), they synchronized significantly their activity in each possible combination of two or even three neurons during short periods of the task. However, no systematic relationship could be detected in relation to movement direction. Hatsopoulos et al. (1998) described in a similar way, by calculating crosscorrelation functions for individual pairs of motor cortical neurons, that synchronous activity did indeed convey directional information, mostly during movement onset. The authors classified, however, a pair of neurons as directionally tuned if synchrony was statistically significant in at least one direction. No systematic relation to movement direction was required as, for instance, determining the preferred direction by fitting a cosine function to the degree of synchronous activity in each movement direction, as directional tuning is most often defined for firing rate (Georgopoulos et al. 1982). In the same way as in Hatsopoulos et al. (1998), no relationship could be detected in our data between the preferred direction determined for the firing rate of each of the neurons in a pair and the highest amount of synchrony, although all pairs of neurons did significantly synchronize their activity during at least one movement direction, most often during more than one. Here we show for the first time, at the population level, a clear systematic relationship between synchronicity and movement direction. Our time-resolved analysis technique revealed directionally tuned synchronicity only during movement preparation at a distinct moment in time toward the end of the preparatory period. The cosine fit of the amount of synchronous activity was statistically significant (Fig. 6c), its directional selectivity index revealed a strong modulation (I ¼ 0:71) and the means of the probabilities for being significantly synchronized for right-sided and left-sided movements were significantly different (Fig. 7a). Although the mean firing rate of the same population of neurons was also significantly tuned to movement direction about 400 ms later (Fig. 6f), during movement onset, its directional selectivity index was rather low (I ¼ 0:11), and the means of individual firing rates during right-sided and left-sided movements were never significantly different during the whole task (Fig. 7b). Interestingly, however, preferred directions of both synchronicity and mean firing rate appeared to be at almost the same movement direction (6.8 vs. 6, Figs. 6c vs. 6f). Recently, Oram et al. (2001) argued that synchronous spikes do not provide more directional information than firing rates. They recorded motor cortical activity during the same task as that presented by Hatsopoulos et al. (1998) but used different analytical procedures. The main differences between Oram et al. (2001) and the present report is that, first, we utilized time-resolved statistics in order to determine and compare the dynamics of both synchrony and firing rate, and, second, our calculations are based on an entire population of neurons and not on individual pairs. The significant directional selectivity of synchrony in our data is based on the number of neurons that significantly synchronized their activity during a given movement direction at a distinct moment in time. We have shown, at the population level, not only that both measures – the mean firing rate and synchronicity – modulated significantly with movement direction, but particularly that modulation strength varied in time and peak values occurred at different epochs during the trial, synchrony largely preceding firing rate. Furthermore, the modulation of both types of activity occurred systematically 371 within an entire population of neurons. Finally, directional modulation appeared to be, at the population level, much larger for synchrony than for firing rate. 4.4 Reaction time Behavioral data revealed that reaction time varied with movement direction in a systematic way, that is, mean reaction times obtained in each movement direction fitted significantly a cosine function (Fig. 1b). The animal’s behavioral preferred direction was close to the most right-sided target, target 1 (PD ¼ 6:7, Fig. 1b). This corresponded almost perfectly to the preferred direction of synchrony at about 300 ms before the occurrence of the response signal (PD ¼ 6:8, Fig. 6c). The question arises whether synchronous neuronal activity is involved in processing the direction of the movement or rather in speeding up its initiation. There is much evidence from the literature that the increase of motor cortical activity during the preparatory period is strongly involved in the reduction of reaction time. It has been shown that not only global EEG activity recorded over human motor cortical areas at the end of a preparatory period (Coles 1989) but also the firing rate of single motor cortical neurons (Kubota and Hamada 1979; Lecas et al. 1986; Riehle and Requin 1989, 1993) were significantly correlated with reaction time. However, as far as we know a correlation between synchronous spiking activity and reaction time has never been described in the literature. In this context, two hypotheses must be discussed: does the increase of synchronicity at the population level represent movement direction per se or does it reflect the behavioral preference of the animal expressed by reaction times? If synchronous spiking activity were involved only in processing movement direction, similar changes in synchronicity in relation to the different movement directions should be observed as in Fig. 5, even when the same reaction times were produced in each of them. In view of the fact that systematically in each behavioral session reaction time varied in relation to movement direction, this hypothesis could not be tested reliably in our data. On the contrary, synchronous activity may increase during trials in specific directions only because of inducing short reaction times and not because of its involvement in processing movement direction. If this hypothesis were true, we should be able to observe similar differences in synchrony in relation to reaction times by dividing each data set (i.e., for each pair of neurons and each selected movement direction) into two equal subsamples of short and long reaction time trials. Unfortunately, the limited number of trials within each data set did not allow for the application of correct statistics for calculating significant synchronization in such subdivisions (see Sect. 2). One argument, however, in favor of the reaction-time hypothesis is that the probability of significant synchronization undeniably depended on reaction time when data were averaged over all movement directions (Fig. 4a) and thus no longer depended on individual movement directions. Indeed, synchronous activity was much lower during the first half of the preparatory period than during the second half. Recall that reaction times after short trials were much longer than after long ones, the highly significant difference of 130 ms (Table 1) being explained by the probability of signal occurrence (p ¼ 0:5 vs. p ¼ 1). The fact that synchronicity, firing rate, and reaction times were tuned to roughly the same movement direction is far from being trivial. First, the directional selectivity of the population firing rate cannot be explained by a directional bias due to the small sample size because the distribution of preferred directions of the selected neurons appeared to be statistically homogeneous (Fig. 1c). Furthermore, it is important to note that the population tuning was calculated by determining the mean activity of all neurons in each time step and each movement direction, irrespective of the preferred directions of the neurons. Since we did not normalize neuronal activity, neurons exhibiting a higher mean firing rate contribute more than neurons with low firing rates. Second, our data and those of others (Hatsopoulos et al. 1998; Grammont and Riehle 1999) did not show any systematic relationship between the directional tuning of the firing rate and significant synchronization. These results lead us to speculate that synchronous activity and firing rate are cooperative neuronal processes and that the directional matching of our three measures – firing rate, synchronicity, and reaction times – might be an effect of behaviorally induced network cooperativity acquired during learning. In other words, the monkey developed during training its behavioral right-sided preference, which shaped the dynamical interplay in a large group of neurons without changing their individual tuning properties. Acknowledgements. We wish to thank Sonja Grün, Markus Diesmann, and Bill MacKay for many helpful and exciting discussions and one anonymous referee for her/his helpful comments. Special thanks go to Annette Bastian for her help in data collection, Michèle Coulmance for writing data acquisition and parts of data analysis software, and Marc Martin for animal welfare. This research was supported in part by the CNRS, GIS (Sciences de la Cognition), and ACI Cognitique (Invariants and Variability). FG was supported by the French government (MENRT). References Abeles M (1982) Role of the cortical neuron: integrator or coincidence detector? Isr J Med Sci 18: 83–92 Abeles M (1991) Corticonics: neural circuits of cerebral cortex. Cambridge University Press, Cambridge Abeles M, Bergman H, Margalit E, Vaadia E (1993) Spatiotemporal firing patterns in the frontal cortex of behaving monkeys. J Neurophysiol 7: 1629–1638 Aertsen A, Diesmann M, Grün S, Arndt M, Gewaltig MO (1995) Coupling dynamics and coincident spiking in cortical neural networks. In: Herrmann HJ, Pöppel E, Wolf DW (eds) Supercomputers in brain research: from tomography to neural network. World Scientific Publisher, Singapore, pp 213– 224 372 Aertsen A, Erb M, Palm G (1994) Dynamics of functional coupling in the cerebral cortex: an attempt at a model-based interpretation. Physica D 75: 103–128 Aertsen A, Vaadia E, Abeles M, Ahissar E, Bergman H, Karmon B, Lavner Y, Margalit E, Nelken I, Rotter S (1991) Neural interactions in the frontal cortex of a behaving monkey: signs of dependence on stimulus context and behavioral state. J Hirnforsch 32: 735–743 Ahissar M, Ahissar E, Bergman H, Vaadia E (1992) Encoding of sound-source location and movement: activity of single neurons and interactions between adjacent neurons in the monkey auditory cortex. J Neurophysiol 67: 203–215 Barlow HB (1972) Single units and sensation: a neuron doctrine for perceptual psychology? Perception 1: 371–394 Bastian A, Riehle A, Erlhagen W, Schöner G (1998) Prior information preshapes the population representation of movement direction in motor cortex. Neuroreport 9: 315–319 Bertelson P (1967) The time course of preparation. Q J Exp Psychol 19: 272–279 Bertelson P, Boons JP (1960) Time uncertainty and choice reaction time. Nature 187: 131–132 Bonnet M, Requin J, Stelmach GE (1982) Specification of direction and extent in motor programming. Bull Psychon Soc 19: 31–34 Coles MGH (1989) Modern mind-brain reading: psychophysiology, physiology, and cognition. Psychophysiology 26: 251–269 deCharms RC, Merzenich MM (1996) Primary cortical representation of sounds by the coordination of action-potential timing. Nature 381: 610–613 Diesmann M, Gewaltig MO, Aertsen A (1999) Stable propagation of synchronous spiking in cortical neural networks. Nature 402: 529–533 Fujii H, Ito H, Aihara K, Ichinose N, Tsukada M (1996) Dynamical cell assembly hypothesis: theoritical possibility of spatio-temporal coding in the cortex. Neural Netw 9: 1303– 1350 Georgopoulos AP, Kalaska JF, Caminiti R, Massey JT (1982) On the relations between the direction of two-dimensional arm movements and cell discharge in primate motor cortex. J Neurosci 2: 1527–1537 Gerstein GL, Bedenbaugh P, Aertsen MH (1989) Neuronal assemblies. IEEE Trans Biomed Eng 36: 4–14 Grammont F (2001) Rôle fonctionnel de la coopérativité neuronale impliquée dans la préparation à l’action. Thèse de Doctorat. Université de Provence, France Grammont F, Riehle A (1999) Precise spike synchronization in monkey motor cortex involved in preparation for movement. Exp Brain Res 128: 118–122 Grammont F, Riehle A (2000) Difference in dynamics of spike synchronization and firing rate in the same population of neurons suggests specific involvement in motor cognitive functions. Eur J Neurosci 12 Suppl 11: 502 Grün S (1996) Unitary joint-events in multiple-neuron spiking activity-detection, significance and interpretation. Harri Deutsch, Thun Grün S, Diesmann M, Aertsen A (2002a) ‘Unitary Events’ in multiple single-neuron activity. I. Detection and significance. Neural Comput 14: 43–80 Grün S, Diesmann M, Aertsen A (2002b) ‘Unitary Events’ in multiple single-neuron activity. II. Non-stationary data. Neural Comput 14: 81–119 Grün S, Diesmann M, Grammont F, Riehle A, Aertsen A (1999) Detecting unitary events without discretization of time. J Neurosci Meth 94: 67–79 Grün S, Riehle A, Diesmann M (2003) Effect of across trial nonstationarity on joint-spike events. Biol Cybern (this issue). DOI 10.1007/s00422-002-0386-2 Hatsopoulos NG, Ojakangas CL, Paninski L, Donoghue JP (1998) Information about movement direction obtained from synchronous activity of motor cortical neurons. Proc Natl Acad Sci USA 95: 15706–15711 Hebb DO (1949) Organization of behavior. Wiley, New York Kubota K, Hamada I (1979) Preparatory activity of monkey pyramidal tract neurons related to quick movement onset during visual tracking performance. Brain Res 168: 435–439 Kurata K (1993) Premotor cortex of monkeys: set- and movementrelated activity reflecting amplitude and direction of wrist movements. J Neurophysiol 69: 187–200 Lecas JC, Requin J, Anger C, Vitton N (1986) Changes in neuronal activity in the monkey precentral cortex during preparation for movement. Neurophysiology 56: 1680–1702 Lépine D, Glencross D, Requin J (1989) Some experimental evidence for and against a parametric conception of movement programming. J Exp Psychol HPP 15: 347–362 Mardia KV (1972) Statistics of directional data. Academic, London. Mountcastle VB, Reitboeck HJ, Poggio GF, Steinmetz MA (1991) Adaptation of the Reitboeck method of multiple microelectrode recording to the neocortex of the waking monkey. J Neurosci Meth 36: 77–84 Nicolelis MA, Baccala LA, Lin RC, Chapin JK (1995) Sensorimotor encoding by synchronous neural ensemble activity at multiple levels of the somatosensory system. Science 268: 1353– 1358 Oram MW, Hatsopoulos NG, Richmond BJ, Donoghue JP (2001) Excess synchrony in motor cortical neurons provides redundant direction information with that from coarse temporal measures. J Neurophysiol 86: 1700–1716 Palm G, Aertsen AM, Gerstein GL (1988) On the significance of correlations among neuronal spike trains. Biol Cybern 59: 1–11 Requin J, Brener J, Ring C (1991) Preparation for action. In Jennings JR, Coles MGH (eds) Handbook for cognitive psychology: central and automatic nervous system approaches. Wiley, New York, pp 357–448 Riehle A, Requin J (1989) Monkey primary motor and premotor cortex: single-cell activity related to prior information about direction and extent of an intended movement. J Neurophysiol 61: 534–549 Riehle A, Requin J (1993) The predictive value for performance speed of preparatory changes in neuronal activity of the monkey motor and premotor cortex. Behav Brain Res 53: 35– 49 Riehle A, Requin J (1995) Neuronal correlates of the specification of movement direction and force in four cortical areas of the monkey. Behav Brain Res 70: 1–13 Riehle A, MacKay WA, Requin J (1994) Are extent and force independent movement parameters? Preparation- and movement-related neuronal activity in the monkey cortex. Exp Brain Res 99: 56–74 Riehle A, Grün S, Diesmann M, Aertsen A (1997) Spike synchronization and rate modulation differentially involved in motor cortical function. Science 278: 1950–1953. Riehle A, Grammont F, Diesmann M, Grün S (2000) Dynamical changes and temporal precision of synchronized spiking activity in monkey motor cortex during movement preparation. J Physiol (Paris) 94: 569–582 Rosenbaum DA (1980) Human movement initiation: specification of arm, direction, and extent. J Exp Psychol Gen 109: 444–474 Roskies AL (ed) (1999) Reviews on the binding problem. Neuron 24: 7–125 Roy A, Steinmetz PN, Niebur E (2000) Rate limitation of unitary event analysis. Neural Comput 13: 2063–2082 Shadlen MN, Newsome WT (1994) Noise, neural codes and cortical organization. Curr Opin Neurobiol 4: 569–579 Singer W (1999) Neuronal synchrony: a versatile code for the definition of relations? Neuron 24: 49–65 Singer W, Gray CM (1995) Visual feature integration and the temporal correlation hypothesis. Annu Rev Neurosci 18: 555– 586 Softky WR (1995) Simple codes versus efficient codes. Curr Opin Neurobiol 5: 239–247 373 Vaadia E, Haalman I, Abeles M, Bergman H, Prut Y, Slovin H, Aertsen A (1995) Dynamics of neuronal interactions in monkey cortex in relation to behavioural events. Nature 373: 515–518 Von der Malsburg C (1981) The correlation theory of brain function. Internal Report 81-2: Abteilung Neurobiologie, MPI für Biophysikalische Chemie, Göttingen Von der Malsburg C (1995) Binding in models of perception and brain function. Curr Opin Neurobiol 5: 520–526 Weinrich M, Wise SP (1982) The premotor cortex of the monkey. J Neurosci 2: 1329–1345 Weinrich M, Wise SP, Mauritz KH (1984) A neurophysiological study of the premotor cortex in the rhesus monkey. Brain 107: 385–414