Optimizing Deep Learning Based on Deep Auto Encoder and Genetic Algorithm

Abstract

Deep learning (DL) approaches have demonstrated the ability to learn useful features directly from data for a wide variety. The difficulties in learning deep generative or discriminative models can be overcome by an initial unsupervised learning step that maps the inputs to useful intermediate representations. DL systems can automatically discover and generate more combination, and high-level features from raw data sources. Auto encoder techniques include Deep Auto encoder (DAE), Stack Auto encoder (SAE), Contractive Auto encoder (CAE), and Denoising Auto encoder (DA). DAE aims to make good representations of data which can be utilized for reconstruction and classification. It is considered as one of the powerful algorithms that gives higher accuracy and best performance. The proposed method in this paper is based on using DAE and Genetic Algorithm (GA) through applying split-training and merging algorithms for DL. First, the network is divided into two initialized networks (DAE1 and DAE2) using DAE then both of these networks are merged using GA with adding additional dataset for Training process in DAE2. This proposed approach was evaluated based on MNIST dataset and the obtained results showed higher accuracy and lower error in the classification.