High capacity neural networks for familiarity discrimination

This paper presents two new novelty discrimination models for uncorrelated patterns based on neural modelling. The first model uses a single neuron with Hebbian learning and works well when the number of memorised patterns is less than 0.046 N (where N is the number of inputs). The second model is based on checking the value of the energy function of a Hopfield network. By sacrificing the ability to extract patterns, not needed for familiarity detection, an amazing jump from the normal capacity for retrieval of 0.145 N to a capacity for novelty discrimination of 0.023 N/sup 2/ is achieved. In addition, both models give some insight on the effect of deja vu, since there is always a very small probability of detecting novel patterns as familiar.