Learning feature constraints in a chaotic neural memory

We consider a neural network memory model that has both nonchaotic and chaotic regimes. The chaotic regime occurs for reduced neural connectivity. We show that it is possible to adapt the dynamics in the chaotic regime, by reinforcement learning, to learn multiple constraints on feature subsets. This results in chaotic pattern generation that is biased to generate the feature patterns that have received responses. Depending on the connectivity, there can be additional memory pulling effects, due to the correlations between the constrained neurons in the feature subsets and the other neurons.