Rapid adaptation enabled by surprise-modulated three-factor plasticity.

A: Error magnitude of the transition matrix (Frobenius norm between the true transition matrix T* and the estimated matrix T) as a function of time for the SpikeSuM model (red), and a Spiking Neural Network model (SNN) with the same architecture and number of neurons as SpikeSuM, but simple modulation (cyan SNN_sm) or no modulation (green SNN_nm), in a volatile sequence task with different stimuli and K = 4 possible transitions. Rule switches cause the occasional abrupt increases in error. The SpikeSuM network exhibits faster learning immediately after the switch as well as better convergence during periods when the rule stays fixed; volatility H=0.001. B Zoom on 200 presentation steps immediately after a rule switch. The red curve goes down faster and to a lower value than the other two. C: The surprise signal transmitted by the 3rd factor as a function of the activity for three cases (red: SpikeSuM rule; cyan: simplified modulation rule; green: constant learning rate, no modulation). The parameters of all three rules have been optimized. D Average error over 10’000 presentation steps with volatility H = 0.001 for different values of (size) and K. The performance of SpikeSuM is comparable to that of the Bayesian Online Change Point detection algorithm (BOCPA, black) and varSMile (grey) and better than SNN_nm or SNN_sm. The results with random connectivity SpikeSuM_rand are shown in dark blue.