Towards a Model of Associative Memory with Learned Distributed Representations

Associative memory is a device capable of storage of data and its retrieval from incomplete or noisy probes. This article describes a neural model of associative memory inspired by continuous Modern Hopfield networks. The proposed learning procedure produces distributed representations of the fragments of input data which collectively represent the stored memory patterns, governed by the activation dynamics of the network. This allows for effective storage of data without the need to grow the network. In comparison to training by error back propagation, the training procedure is relatively fast (few-shot learning).