Adaptive Learning Rate for Unsupervised Learning of Deep Neural Networks

In this paper an approach for adaptive learning step calculation using ReLU transfer function in neural network is proposed. This adaptive learning rate aims to automatically choose the step size that minimizes the objective function of neural network. We give a theoretical justification for the proposed adaptive learning rate approach, which is based on the steepest descent method. The main contribution of this paper is a novel technique for adaptive learning rate calculation, if we use ReLU transfer function. The experiments in data compression datasets show that proposed approach provides better generalization capability (test set accuracy) and permits to choose the learning rate automatically.