An internal mechanism for detecting parasite attractors in a Hopfield network

This paper presents a built-in mechanism for automatic detection of prototypes (as opposed to parasite attractors) in a Hopfield network. It has a good statistical performance and avoids host computer overhead for this purpose. This mechanism is based on an internal coding of the prototypes during learning, using cyclic redundancy codes, and leads to an efficient implementation in VLSI. The immediate separation of prototypes from parasite attractors can be used to enhance the autonomy of the networks, aiming at hierarchical multinetwork structures. As an example, the use of such an architecture for the classification of handwritten digits is described.