LEARNING METHOD OF MULTI-LAYER PERCEPTRONS WITH N-BIT DATA PRECISION

摘要

The present invention relates to a method of learning by N-bit data representation of a multilayer perceptron neural network. In the conventional digital learning, large numbers and small numbers cause underflow and overflow by quantification. In order to solve the problem that the size of data representation bits had to be large due to the constraints of bit truncation, the weighted sum calculation has 2N bit data precision in the forward and backward calculations in the N-bit digital learning of the multilayer perceptron. When expressing the weighted sum result by 2N-bit data precision as N-bit data for sigmoid nonlinear transformation in omni-computation, set the maximum value of N-bit representation to the value corresponding to the saturation region of sigmoid, In the perceptron reverse calculation, the weighted sum result by 2N bit data precision is N. When expressed in bit data, the maximum value of the N-bit representation is set to be relatively smaller than the maximum value represented by 2N bits, the weight expression range at the beginning of the learning is reduced, and the weighting ratio is reached when the constant ratio reaches the maximum value as the learning progresses. By performing a learning method that extends the range of, the 8-bit digital learning performance can be improved by 16-bit digital learning performance.