Generalized Hopfield networks for associative memories with multi-valued stable states

Hopfield networks with fully connected standard neurons can be generalized by replacing bi-level activation functions with their multilevel counterparts. Multilevel neuron characteristics are discussed in the paper with emphasis on their inflection points. It is shown that an activation function possessing (N + 1)-levels yields N + 1 minima and N saddle points of the computational energy function when two generalized neurons are used in a conventional bi-stable connection. Analytical results for parameter constraints and energy function properties are discussed for binary and ternary characteristics of neurons. Gradient fields indicating basins of attraction for continuous-time networks are used to illustrate dynamical relationships during network convergence to stable points. Results indicate that generalized Hopfield networks can be used for multilevel signal processing and smoothing of planar images.

[1]  Jacek M. Zurada,et al.  Stationary points of single-layer feedback neural networks , 1992, [Proceedings] 1992 IEEE International Symposium on Circuits and Systems.

[2]  Robert W. Newcomb,et al.  A multilevel neural network for A/D conversion , 1993, IEEE Trans. Neural Networks.

[3]  Bing J. Sheu,et al.  Modified Hopfield neural networks for retrieving the optimal solution , 1991, IEEE Trans. Neural Networks.

[4]  J. Hopfield,et al.  Computing with neural circuits: a model. , 1986, Science.

[5]  Anthony N. Michel,et al.  Analysis and synthesis of discrete-time neural networks with multilevel threshold functions , 1991, 1991., IEEE International Sympoisum on Circuits and Systems.

[6]  Peter Aronhime,et al.  Vector field analysis of single layer feedback neural networks , 1990, Proceedings of the 33rd Midwest Symposium on Circuits and Systems.

[7]  Y Ichioka,et al.  Hopfield model with multistate neurons and its optoelectronic implementation. , 1991, Applied optics.

[8]  J J Hopfield,et al.  Neurons with graded response have collective computational properties like those of two-state neurons. , 1984, Proceedings of the National Academy of Sciences of the United States of America.

[9]  Jacek M. Zurada,et al.  Introduction to artificial neural systems , 1992 .