Computation of Weighted Sum by Physical Wave Properties — Coding Problems by Unit Positions
An architecture for neural computation is proposed. The costliest parts of neural computation: inter-unit-communication, weight representation, and weighted sum execution are all implemented by natures of wave propagation and interaction. As a result, unit interaction is realized without wires, weighted sum is executed without adders or multipliers, and the weight values are coded as unit positions. Some experimental results, which are based on software models of the wave natures, show that the approach provides a promising computation potential regardless of its limited ability to represent connection weights.
KeywordsRecurrent Neural Network Output Unit Neural Computation Weighting Mechanism Unit Position
Unable to display preview. Download preview PDF.
- L.B. Almeida. Backpropagation in perceptions with feedback. Neural Computers, 1988.Google Scholar
- J.J. Hopefield and D.W. Tank. Neural computation in optimization problems. Biol. Cybernetics, 52, 1985.Google Scholar
- I. Kumazawa. A signal sharing scheme for unifying computation and learning phases of recurrent networks. In Artificial Neural Network II Elsevier, 1992.Google Scholar
- I. Kumazawa. Neural computation by interaction of waves: Coding problems by unit positions. Technical report, Neural Computation Group, March 1997. Just submitted.Google Scholar
- I. Kumazawa and Y. Kure. Computation by interaction waves: Realisation of recurrent neural networks. In IEICE, Spring National Conference, 1997. Just submitted.Google Scholar
- Y. Kure and I. Kumazawa. Neural computation by interaction of waves: Weight representation by differences in wave phases. Technical report, Neural Computation Group, March 1997. Just submitted.Google Scholar
- F.J. Pineda. Generalization of back-propagation to recurrent neural networks. Physical Review Letters, 59, 1987.Google Scholar