Figure 4. A computational model of similarity-based generalization.
(A) Schematic of the model. When a stimulus orientation xk is presented, the predicted reward for this orientation Vk is computed by integrating the excitatory and inhibitory associations E and I of all stimuli j that are similar to k, (including its own associations, j = k), weighted by the similarity between stimuli j and k. The similarity between j and k is determined by the excitatory and inhibitory generalization coefficients eSjk and iSjk, respectively, which are assumed to be Gaussian (or exponential, not shown here). The width of the excitatory and inhibitory generalization coefficients, and thus the degree to which excitatory and inhibitory associations generalize from j to k is determined by the parameters se and si. The reward prediction Vk is used to generate approach behavior P(+) and to compute a reward prediction error δ, which in turn updates the excitatory and inhibitory associations of k. (B) Illustration of the effects of changes in the width of excitatory and inhibitory generalization coefficients on generalization gradients.