, 2010), presumably via a complex web of local interconnections. Decision making in this region of prefrontal cortex is therefore best characterized as a transition from a context-invariant state to context-dependent

coding within the same functional network. Our results are consistent with an adaptive coding model of prefrontal cortex in which flexible goal-oriented behavior is mediated via dynamic changes to prefrontal tuning properties (Duncan, 2001; Duncan and Miller, 2002). As in previous studies (e.g., Freedman et al., 2001; Meyers et al., 2012; Watanabe, 1986), we show that PFC processes input as a function of task relevance. Here we provide a detailed picture of the underlying network dynamics, from rule encoding and maintenance to context-dependent decision making. A plausible mechanism for flexible tuning is activity-dependent short-term synaptic plasticity (Zucker and

Regehr, 2002). Short-term plasticity BAY 73-4506 has recently been identified as a possible basis for maintaining information in WM (Erickson et al., 2010; Fujisawa et al., 2008; Mongillo et al., 2008). If patterned activity leaves behind a patterned change in the Palbociclib synaptic weights of the network (i.e., hidden state), then subsequent stimulation will be patterned according to the recent stimulation history of a network (Buonomano and Maass, 2009). Thus, any driving input to the system will trigger a systematic population response that could be used to RNASEH2A decode the recent stimulation history of the network (Nikolić et al., 2009). Exactly this phenomenon is seen in our data during the presentation of the neutral stimulus (Figure 5). Although this stimulus was fixed across trials, the population response was patterned according

to the identity of the previous cue, providing a more reliable readout of the memory content than the population response observed during the relatively quiescent delay period. Recent WM studies in human (Lewis-Peacock et al., 2012) and nonhuman primate (Barak et al., 2010) have also proposed a similar mechanism for maintaining the contents of WM. Short-term synaptic dynamics could also explain nonstationary population activity profiles, as observed here (Figure 4) and in other studies (Barak et al., 2010; Crowe et al., 2010; Meyers et al., 2008). If the hidden state of the network is continually altered by each pattern of activity, then even constant input to the system should result in time-varying patterns (Buonomano and Maass, 2009). Indeed, it could be relatively difficult to engineer a network that maintains a static activity state in the presence of activity-dependent short-term plasticity. Finally, adaptive changes in tuning mediated by short-term synaptic dynamics could also explain the differential activity states observed during the delay period. Analysis of the neutral stimulus suggests that differences in the underlying hidden state can be revealed by increasing overall activity in the network.