ABSTRACT
We describe a public-domain simulator specialized for discrete asynchronous Hopfield models. There is a careful efficient sequential implementation of asynchronous dynamics that is also well-suited to vector machines (the algorithm was implemented also on the Connection Machine CM-5 to confirm this). Any asynchronous dynamics and any on-line associative memory storage rule are supported. A symbolic input/output interface is provided: units are named; patterns are sets of ON units. Hashing is used for efficient symbol-to-unit mapping. The network grows during the training phase: units are automatically created for previously unseen symbols in the input. There is no inherent limit on the size of the network that may be created. The simulator has been instantiated for a special case of the discrete asynchronous Hopfield network proposed by the author. Over the last three years, it has been applied to associative memory problems (machine-printed word recognition; character restoration), knowledge representation, and combinatorial optimization (Approximating MAX-CLIQUE; Constraint Satisfaction Problems).
