Pattern-Recognition Methods Applied to the Problem of Noise-Proof Decoding

摘要

The problem of noise-proof decoding is considered as a particular case of pattern recognition. The use of a perceptron as a decoder in a binary symmetric channel is substantiated. An algorithm is developed which generalizes the nearby elements of the input space and which can be realized by threshold elements. A system composed along these lines was simulated on an Ural-2 digital computer. With a constant norm of output vectors, the mathematical expectation of the decision-function argument, in the elementary perceptron, behaves like a decreasing function of the code distance; hence, the perceptron is suitable for operation as a decoder in a temperature-receiver system. Sixteen standard code vectors were selected for the verification of noise rejection of an experimental recognition system. These vectors were used for teaching the simulated (on the computer) perceptron. The perceptron structure (parameters, ties) may be optimized to solve a concrete problem. Thus, a tentative scheme composed of dynamic neurons seems promising for correcting multiple random independent errors.