Automatic Hyperparameter Tuning in Deep Convolutional Neural Networks Using Asynchronous Reinforcement Learning

Major gains have been made in recent years in object recognition due to advances in deep neural networks. One struggle with deep learning, however, revolves around the fact that currently it is unknown what network architecture is best for a given problem. Consequently, different configurations are tried until one is identified that gives acceptable results. This paper proposes an asynchronous reinforcement learning algorithm that finds an optimal network configuration by automatically adjusting parameters for a given problem. It is shown that asynchronous reinforcement learning is able to converge on an optimal solution for the MNIST data set.