The Improvement of Dropout Strategy Based on Two Evolutionary Algorithms

Abstract

Dropout strategy is a simple and common regularization method in the construction of deep network that it can control the status of units in the Dropout layers according to the constant probability values in the training processes to prevent the training from overfitting. However, the probability values of the Dropout strategy are single and decided by users, which means that we need more training iterations to receive better results and avoid less fitting problem. In this paper, two evolutionary algorithms, genetic algorithm and differential evolution algorithm are used to optimize the set probability values of network units to improve dropout strategy and they are proved to be able to increase the accuracy of the original method to about 5%.