Parallelization of Deep Convolutional Neural Network

Abstract

Convolutional Neural Network (CNN) in Deep Learning (DL) has been has been achieving success in various objectives of pattern recognizing and classification. Training CNN model in an acceptable time is necessary. This is computationally intensive task. We have analyzed speedup in training of complex DL model using single core CPU and GPU. In the distributed setting, we studied weight update algorithms. We analyzed performance characteristics using our own designed DL model for facial expression recognition task. The novelty of this effort is to demonstrate performance acceleration in distributed DL frame-work. Analysis and study show that DL training performance improved over six times speedup for the own model.