Performance issues of parallel, scalable convolutional neural networks in deep learning

Abstract

In this work, we investigate the performance issues in the parallel and scalable of Convolutional Neural Networks (CNNs). This will accelerate the training performance of CNN. In this paper we propose the parallel recognition using Compute Unified Device Architecture (CUDA) Technology and Message Passing Interface (MPI). We demonstrate scalability and performance that can be achieved on the GPU using CUDA framework where the computation-intensive tasks shifted on GPU. It compares result on GPU hardware architecture with the serial algorithm on CPU. The main novelty of our method is a new scalable CNN architecture that integrates a category hierarchy with deep CNN.