Cellular-Neural Associative Memory: limiting capability.

List of the participants:
Summary:
The VLSI implementation of Artificial Neural Networks is limited by the necessity of implement the full-graph interneuron communication. The limitation may be removed if the interconnection structure is reduced to a cellular structure with a certain size of neuron neighborhood. Up till now there is no answer what are the directions of searching the ways to compensate the decrease of interneuron communication so, that the space-time characteristics of the algorithms remain acceptable.

The aim of the project is to develop a formal model of cellular-neural algorithm, which is the combination of cellular automaton connection structure with neuron networks connectionist character of computation. Based on such a model some investigations of cellular-neural associative memory of Hopfield type are to be performed. The goal of the investigations is to determine limiting storing capacity and restoring capability in relation to the size of the neuron neighborhood and the properties of stored patterns.

Following investigations are intended to be done:
  1. To perform the comparative analysis of existing methods of Hopfield Associative memory learning from the point of view of the compatibility to cellular nature of interconnections, and chose the most suitable principle for cellular case adaptation.
  2. To modify the chosen method for the cellular case, prove its correctness and test it by simulating.
  3. To study the properties of stability and degree of attraction of stored patterns both theoretically (obtaining necessary and sufficient conditions of individual and asymptotical stability) and experimentally by simulation.
  4. To develop a method for cellular neural algorithm synthesys at given technical and qualitative restrictions.
List of publications:
  1. O.Bandman Cellular-Neural Computations. Formal Model and Possible Applications // Lecture Notes in Computer Science, 964, 1995, p.21-35.
  2. O.Bandman. Cellular-Neural Computation. Formal Model and Possible Applications // Bulletin of the Novosibirsk Computing Center, series: Computer Science, issue 2 (1994). - P.25-44.
  3. O.Bandman. Cellular-Neural Algorithms (methods of representation and computer simulation) // Proceedings of the Conference "New information technologies in discrete structures study". Ekaterinburg, 1996.- P.162-168.
  4. O.L.Bandman, S.G.Pudov. Stability of stored patterns in cellular-neural associative memory.// Bulletin of the Novosibirsk Computer Center. Series: Computer Science, issue 4 (1996), pp.1-16.
  5. Olga Bandman, Sergey Pudov. Design and Simulations of Cellular Neural-like Associative Memory. // IEEE Proceedings of First International Workshop on Distributed Interactive Simulation and Real Time Applications (DIS-RTA'97), January 9-10  Eilat, Israel, pp.49-56.
  6. S.G.Pudov. Learning of Cellular-Neural Associative Memory.  // Avtometriya, N2, 1997, pp.107-120.
  7. S.G.Pudov. Influence of Self-Connection Weights on Cellular-Neural Network Stability. // Lecture Notes in Computer Science, 1277, pp. 76-82.


Please contact Dr. Olga Bandman for all the questions concerning this project.
E-mail: bandman@ssd.sscc.ru


Last update: October 22, 1999