The Hopfield Model and Beyond.
Abstract
The standard Hopfield model (both digital analog) and algorithms to improve its performance are reviewed. An analysis of the model and the modification algorithms is given. Future directions for continuous models which have both large capacity and good error-correcting capabilities are examined. In 1982, Hopfield proposed a neutral model of memory storage and retrieval based on the theory of spin glasses in solid state physics. In the model, neurons are binary-valued threshold units, taking either the value 0 or 1 in one version, or 1 or -1 in an alternative version. This digital restriction of the neurons represents the neuron in two possible states-a 1 represents a neuron that is firing, while a 0 or a -1, a neuron that is inactive. Mathematically, this corresponds to replacing the experimentally observed neuronal input-output relationship, a graded response which can be characterized by a sigmoid function, with a step-function. The neurons form a single layer and are completely interconnected, with the strength of these connections, or synapses, given by a correlation matrix formed from the memory states to be stored in the system.
Document Details
- Document Type
- Technical Report
- Publication Date
- May 15, 1987
- Accession Number
- ADA180607
Entities
People
- Charles M. Bachmann
Organizations
- Brown University