Figure 2. Task-set driven behavior in the network model with an idealized, perfect encoding of task-sets.

The behavior of the model is compared in presence (red lines) and in absence (blue lines) of the inference signal from the task-set network, that allows task-set retrieval. (a,d,g) Model dynamics following an episode switch (at trial zero, the correct task-set shifts without explicit indication). (a) Strengths of synapses in the associative network between neural populations representing the new task-set (solid lines) and the previous task-set (dashed lines). (d) Performance (proportion of correct responses). (g) Mean change $<J_{INC}\cdot (1-J^{AN})>$ in the AN synaptic weights due to the inference signal from the TN. Here the inference strength $J_{INC}$ is one, so that the weights in the AN reach their maximal values as soon as the network makes a first correct choice, and do not change afterwards. The first correct choice takes place randomly on different trials in different episodes leading to a spread over the first trials in the episodes and vanishing changes towards the end of the episode. (b,e,h) Task-set retrieval: same quantities as in (a,d,g), but aligned at the time of the first correct response. (c,f,i) Effect of misleading feedback: same quantities as in (a,d,g), aligned on a misleadingly non-rewarded correct trial at the end of episodes. Average of 5000 sessions of 25 episodes, with 10% of noisy trials. Network parameter values: $\alpha =0.4$, $\beta =7$, $ϵ=0$, $J_{INC}=1$.