ON THE QUESTION OF DIGITAL PERCEPTRON SYNTHESIS (K VOPROSU SINTEZA DISKRETNOGO PERSEPTRONA),
Abstract
The article presents a solution to the problem of synthesizing the logical net of a digital perceptron based on optimum coding and linear programming methods. The structural circuit utilizes simultaneously both recurrent and nonrecurrent forced learning procedures. The recurrent procedure for learning to recognize objects (presented by binary numbers) is based on the simplex method of linear programming. The nonrecurrent procedure is based on the Browning-Bledsoe method of binary digital recognition. Information signs of recognized forms are extracted by computer procedures in general coding (in the sense of the Browning-Bledsoe method), leading to an investigation of the fine structure of the binary numbers by methods in the theory of optimum coding. The algorithms of optimum coding of binary numbers are based on a comparison of divergence and entropy of a series of possible states of the recognition device. Optimum selections of the weighted coefficients in the logical net are determined on the basis of algorithms of the simplex method using digital computers. The authors conclude that the further development of the theory of discrete (digital) perceptron depends on perfecting the recurrent and nonrecurrent learning procedures in pattern recognition devices which involves the use of more general computational methods of mathematical programming. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Aug 04, 1967
- Accession Number
- AD0669270
Entities
People
- E. N. Perevezentsev
- K. N. Denisov
Organizations
- National Air and Space Intelligence Center