Adequate Input for Learning in Attractor Neural Networks

Abstract

In the context of learning in attractor neural networks (ANN) we discuss the issue of the constraints imposed by the requirement that the afferents arriving at the neurons in the attractor network from the stimulus, compete successfully with the afferents generated by the recurrent activity inside the network. We simulate and analyze a two component network: one representing the stimulus, the other an ANN. We show that if stimuli are correlated with the receptive fields of neurons in the ANN, and are of sufficient contrast, the stimulus can provide the necessary information to the recurrent network to allow learning new stimuli, even in very disfavored situation of synaptic predominance in the recurrent part.

On leave of absence from Racah Institute of Physics