An effective Reinforcement Learning method for preventing the overfitting of Convolutional Neural Networks

Convolutional Neural Networks are machine learning models that have proven abilities in many variants of tasks. This powerful model sometimes suffers from overfitting. paper proposes a method based on Reinforcement Learning for addressing this problem. In research, the parameters target layer Network take as state Agent section. Then gives some actions forming hyperbolic secant function. function’s form is changed gradually and implicitly by proposed method. The inputs function weights layer, its outputs multiply same to updating them. study, inspired Deep Deterministic Policy Gradient because into continuous domain. To show method’s effectiveness, classification task considered using Networks. 7 datasets been used evaluating model; MNIST, Extended small-notMNIST, Fashion-MNIST, sign language CIFAR-10, CIFAR-100.