(a-e) Learning a new task-set that partially overlaps a previously learned one. (a) The model is simulated for the three non-overlapping task-sets of the recurrent session for episode 1 to 25, as in Figure 3a,c,e,g. After episode 25, we introduce a fourth task-set that has one overlapping stimulus-action association with task-set 1 (association [5 j]). (b) Performance after the first correct trial for association [5 j] of each episode. Trials were classified depending on whether the episode was between number 25 and 30 (‘during transition’), or after 35 (‘after transition’). As expected, the model predicts a lower performance during the transition, because of incorrect inference of task-set one from the TN to the AN. (c) Average values of task-set network synaptic strengths between neural populations corresponding to each of the three correct task-sets, excluding the strengths from the neural population selective to the overlapping association [5 j]. Task-set four is learned in the task-set network from episode 26. (d) Average value of the task-set network synaptic weights from the neural population selective to the overlapping association [5 j], to neural populations selective to non-overlapping associations, for task-set 1 (in blue) and task-set 3 (in red). Pre-activated depression from the overlapping association [5 j] implicated in the learning of task-set four produces unlearning of connexions between [5 j] and correct associations of task-set 1 ([1d] and [3 f]), thus unlearning of task-set 1. (e) Average value of the inference signal from the task-set network to the associative network connectivity. This inference signal is represented in blue for task-set 1, and in red for task-set 4. At episode 26, the inference is initially incorrect (inference towards task-set one each time the overlapping association [5 j] is activated in the task-set network). As weights for task-set one go below the threshold , this incorrect inference goes to zero. (f,g) Learning three task-sets when two overlap from the beginning. (f) Task-set one and task-set three are overlapping (association [5 j]). (g) Performance after the first correct trial of each episode, in all episodes where either task-set one or task-set three were correct. Trials were classified depending on whether the correct association made was the overlapping [5 j] or a non-overlapping one (‘independent’). As expected, the model predicts a lower performance after the overlapping association, because of incorrect inference of both task-sets from the TN to the AN. Task-sets presentation is periodic for illustration purposes. The simulation corresponds to the average over 500 runs of the recurrent session. Parameters values are , , , and . For clarity, we did not introduce tricky trials in these simulations.