RESEARCH ON BIAX TYPE ELEMENTS AND ASSOCIATED CIRCUITS (BIAX PERCEPTRON)
Abstract
Based upon the experimental work reported, the following three principles, which appear to be fundamental to successful multi-element network learning, have been formulated: (1) excess network capacity, (2) network self- evaluation, and (3) least-effort adaptation. A consistent body of experimental evidence supporting these principles has been developed. A parallel logic technique has been developed which makes possible the economical mechanization of large learning networks. In order to provide a standard of comparison for multi-element learning networks, a listing of all possible linear-input logic mechanizations of functions through four input variables was prepared. Based upon this list, functions can be arranged in some order of relative learning difficulty. The function difficulty, and possible forms of mechanization, therefore represents a standard against which experimental results on specific functions and learning methods may be compared. The list of four-input functions is presented in the Appendix. Experiments on learning in multi-element networks were conducted on the learning machine previously constructed. Based upon these experiments, and other simulations, new learning network principles were formulated. The new principles of network learning were reduced to specific form and hand simulated. The principles may be brief ly described as: (1) excess network capacity, (2) network self-evaluation, and (3) least-effort adaptation.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jan 01, 1963
- Accession Number
- AD0404048
Entities
People
- C. J. Munsey
- J. K. Hawkins
- R. A. Stafford