Hopfield networks as a model of prototype-based category learning: A method to distinguish trained, spurious, and prototypical attractors

We present an investigation of the potential use of Hopfield networks to learn neurally plausible, distributed representations of category prototypes. Hopfield networks are dynamical models of autoassociative memory which learn to recreate a set of input states from any given starting state. These networks, however, will almost always learn states which were not presented during training, so called spurious states. Historically, spurious states have been an undesirable side-effect of training a Hopfield network and there has been much research into detecting and discarding these unwanted states. However, we suggest that some of these states may represent useful information, namely states which represent prototypes of the categories instantiated in the network's training data. It would be desirable for a memory system trained on multiple instance tokens of a category to extract a representation of the category prototype. We present an investigation showing that Hopfield networks are in fact capable of learning category prototypes as strong, stable, attractors without being explicitly trained on them. We also expand on previous research into the detection of spurious states in order to show that it is possible to distinguish between trained, spurious, and prototypical attractors.