Fig. 1. The forgetting curve for a single neuron with supervised learning.

a A sensorimotor circuit model in which cortex and thalamus contain state information (e.g., cues and sensory feedback) and provide input to striatum, which learns to drive actions that lead to reward. b A model of a single striatal neuron receiving cortical inputs, where weights w are trained such that random input patterns x^μ produce the correct classifications $z^{\mu }={\widehat{z}}^{\mu }$. c The input vector x^μ defines a hyperplane (red line) in the space of weights w, and the update rule modifies the weight vector w^μ to give the correct classification of pattern μ with margin κ = 1 (dashed red line). d The probability of incorrect classification when testing pattern ν after learning P = 2N_x patterns sequentially. The most recently learned patterns are on the right, with earlier patterns on the left. The solid curve is the theoretical result; points are simulated results. Inset shows the same result for a perceptron trained with P =  5N_x patterns.