A Novel Method to Fix Numbers of Hidden Neurons in Deep Neural Networks

In this paper, we propose a novel method which can automatically find the preferable hidden neuron numbers in deep neural networks. This method is completed by two cooperating algorithms: Principle Components Analysis (PCA) and Reinforcement Learning (RL). PCA is used to find a range of hidden neuron numbers, and RL is applied to search a better number of hidden neurons and update the searching points. The training process is layer wisely conducted and finally formed a deep neural network. Testing on the MNIST dataset shows, the algorithm can automatically fix the number of hidden neurons layer wisely in deep neural networks and achieve an accuracy of 98.24%, which shows that our method is effective in selection of hidden neuron numbers.