A parallel correlation learning by threshold fluctuation for multilayered perceptrons

Autor: Mitsuaki Mitani, Takahumi Oohori, Kazuhisa Watanabe, Keizo Nishida
Rok vydání: 1996
Předmět:
Zdroj: Systems and Computers in Japan. 27:37-45
ISSN: 1520-684X
0882-1666
Popis: This paper proposes the parallel correlation learning, which is a method different from the conventional back-propagation method. The proposed learning method is applicable to the feedforward-type network in which the output of each unit is differentiable in regard to the internal activation and the output error function is partially differentiable in regard to the set of outputs from the visible units. For the purpose of learning, the time-fluctuation nk(t) (where k is the unit number), which is of small amplitude, unbiased, independent and of uniform distribution, is introduced into the threshold of each unit. The error function value also contains the time-fluctuation due to nk(t), and is represented as e(p, t) (where p is the learning sample number). The correction for the connection coefficient Wk is represented as ΔpWk = −ee(p, t)nk(tqk(p, t), based only on the correlation e(p, t)nk(t) between e(p, t) and nk(t), as well as the average output qk(p, t) of the set of units connected directly to the unit k. It is shown theoretically that the steepest descent of the error is achieved for the state where all fluctuation components are suppressed. In the proposed learning method, all connection coefficients are trained in parallel independently of the shapes of the error function or the nonlinear function of the unit without considering the layer structure. Through the simulation for XOR problem, the learning performance is verified.
Databáze: OpenAIRE
Externí odkaz: