Matteo Negri — Università di Roma Sapienza #
Random Feature Hopfield Networks generalize retrieval to previously unseen examples #
It has been recently shown that, when an Hopfield Network stores examples generated as superposition of random features, new attractors appear in the model corresponding to such features. In this work we show that the network also develops attractors corresponding to previously unseen examples generated with the same set of features. We claim that this surprising behaviour is due to the formation of attractors in correspondence of mixtures of features, and we support this claim by calculating analytically the phase diagram. Finally, we discuss how this framework could be used to predict generalization capabilities of modern neural networks.