Supervised learning with artificial selection

Abstract

Summary form only given, as follows. Supervised learning with artificial selection is proposed as a way to escape from local minima. The concept of artificial selection is reasonable for nature. In the authors' method, the 'worst' hidden unit is detected, and then all the weights connected to the detected hidden unit are reset to small random values. According to simulations, only half the trials using conventional backpropagation converge, whereas all of the trials using the proposed method converge, and quickly do so.<>