Fig. 2. Repetition during training improves recall in the two-pathway model.

a The two-pathway architecture, with cortical and thalamic inputs to striatum (top), where fast supervised learning of corticostriatal weights w is accompanied by slow Hebbian learning of thalamostriatal weights v (bottom). b The forgetting curve (top) in a case where six patterns are repeated multiple times during training, while all other patterns are presented only once (bottom). Solid curve is theoretical result (α = 1, β = 1); points are simulations with N_x = N_y = 1000; dotted curve shows the case in which no patterns are repeated multiple times. c The error rate for pattern ν*, which is repeated $n_{{\nu }^{*}}$ times during training, while other patterns are presented only once. d The error rate as a function of $n_{{\nu }^{*}}$, with curves corresponding to different choices of ν*. e Top: A single pattern ν* is trained multiple times while all other patterns are trained once. Bottom: the number of times that pattern ν* must be repeated during training in order to obtain a classification error rate below a threshold p_θ during testing.