Showing papers by "Konstantinos G. Margaritis published in 1995"

On the systolic implementation of associative memory artificial neural networks

Abstract: This paper describes the systolic implementation of a class of Artificial Neural Networks generally termed as Associative Memories. Such networks are the Discrete Autocorrelator or Discrete Hopfield, the Bi-directional Associative Memory, the Temporal Associative Memory, the Linear Associative Memory and the Novelty Filter. Further, problems such as different weight updating strategies, higher order correlation techniques and Fuzzy pattern encoding are discussed. The Re-Usable Matrix Vector and Matrix Multiplication Systolic Arrays form the basis of the systolic implementation.

Folding techniques for systolic iterations

Abstract: This paper discusses the application of folding techniques on systolic matrix vector iterations performed on re-usable linear systolic arrays. The matrix vector iterations are presented in the context of well known linear algebra iterative methods for the solution of systems of equations. The same iterations form the basis of relaxation procedures of Hopfield type artificial neural networks.

On multi-property set operations using neural networks and systolic arrays

Abstract: This paper presents artificial neural networks and systolic architectures that detect multi-property relations satisfied by the elements of a se! that is: elements having some common properties; satisfying at least one of some given properties; not satisfying all given properties; not satisfying any of some properties. The operations discussed are initially performed by means of a feedforward artificial neural network which encodes in its interconnection matrices the values of the properties of the set members. Then the systolic implementation of the neural network is addressed. Finally the ability of modifying, adding or deleting elements or properties is discussed. Keywordg neural networks, systolic algorithms, set operations