Deterministic transitions have a different signature than stochastic ones.

The paradigm uses a volatile sequence task with a re-occurrence of rules but restricted to different (auditory or visual) stimuli and two different transition rules (A and B) and could be tested in rodent experiments.A Transition matrix corresponding to rule A (left) and rule B (right). The transition to stimulus ‘0’, T_7→0 = 1, is deterministic (yellow square, lower left corner) under rule A but stochastic with a value of T_4→0 = 0.5 (light blue) under rule B, and vice versa for the transitions stimulus ‘4’.B Population activity averaged over network neurons in populations P1 and P2 during all presentations of stimulus x_t+1 = 0 (left) or x_t+1 = 4 (right). Black lines: SpikeSuM without context. Orange/blue lines: population of neurons in module 1 of SpikeSuM-C. Green/red lines: population of neurons in module 2 of SpikeSuM-C. Module 1 learns to implement rule A (indicated by decreased activity). Horizontal axis: count of occurrences of stimulus ‘0’ or ‘4’, respectively. Inset, middle: histogram of average activity after 200 presentation time steps under a given rule. Black bars: comparison of activity under rules A and B in SpikeSuM without context. Colored bars: The activity of neurons in module 1 of SpikeSuM-C during stimuli under rule A is compared with that of neurons in module 2 during stimuli under rule B. In both cases, the network is driven by the same stimulus, but a stochastic transition causes more activity than a deterministic one, since the prediction in the stochastic setting is less reliable.C, same as in B, but only the activity of those neurons responsive to stimulus ‘0’ (left) or ‘4’ (right) is shown. In contrast to the simple SpikeSuM network without context, neurons in the SpikeSuM-C network that respond to stimulus ‘4’ in module 1 under rule A (blue) respond even stronger in the context of rule B but this does not affect their plasticity. Thus the same experimental paradigm also differentiates between models with and without context modules.

(EPS)