This suggests that the low-energy state in the delay period is also stable across time. Importantly, this velocity metric is sensitive to changes in the state of the network, even if the overall energy of the system remains constant. Therefore, multidimensional velocity provides a richer measure of the population dynamics than overall change in activity levels (shown in Figure 2E, bottom), which reveals only a single dominant peak at around 85 ms corresponding to the initial increase in firing at stimulus onset, followed by a second smaller increase in energy

change at around 250–300 ms that tracks the gradual decrease in firing rate observed across the population. Overall, these initial analyses show that the transient increase

in neural firing triggered by the instruction cue is associated with a rapid configuration C646 of activity in state space that differentiates trial type. Activity then settles into a relatively low-energy stable state toward the offset of the cue and into the delay period. Although separation by trial type becomes less distinct during this more quiescent phase, the population response remains statistically separable. To explore the dynamic evolution of activity states discriminating different trial types, we exploited a cross-temporal selleck products variant of pattern classification (see schematic in Figure 3A). First, we demonstrate that the general classification approach is able to decode information content from the pattern of activity observed after the cue presentation. This time-resolved pattern analysis demonstrates significant coding of the cue at around 100 ms (Figure 3B), corresponding to the time of rapid divergence observed in the distance metric (Figure 2B). Pattern classification also peaks at around 230 ms and remains relatively uniform into the delay period. To directly assess the time stability of the activity state differentiating trial types, we decoupled the temporal windows used

for train and test (see schematic in Figure 3A; see also Crowe et al., 2010; Meyers et al., 2008). If accurate generalization is observed across time (train at time t, test at time t+n), we can infer that the population code that differentiates trial type at time t is significantly similar to the coding scheme at time t+n. At the extreme, if the coding schemes were completely Tryptophan synthase time stable, pattern classification should not be sensitive to which time points are used for test or train—by definition a stationary code does not vary across time. Conversely, if classifiers trained at time t are unable to decode patterns observed at time t+n, then we can conclude that population coding is time specific. Cross-temporal classification results for trial type are presented in Figure 4. Different color traces represent classification performance for classifiers trained on data from corresponding shaded time windows and tested throughout the cue and delay epochs.