Proposal of D-DALP: A Method to Automatically Tune the Number of Hidden Units Effective for Deep Learning

Abstract

Although deep learning is an effective method to conduct a machine learning, there is no definite answer on how to initialize the hyper parameters. In this paper, we propose an extension of DALP algorithm, an algorithm to automatically tune the number of hidden units for RBM, Double DALP (D- DALP). D-DALP is an algorithm that automatically tunes the number of hidden units that is effective for deep learning by applying DALP twice to the same data set. In the experiment, it is shown that RBM initialized with D-DALP has a higher identification accuracy when compared to RBM initialized with DALP in a deep learning setting.