Steps in deployment and development of Convolutional Neural Network based applications

Abstract

Designated among 10 breakthrough technologies by MIT Technology Review [1], Deep Learning (DL) outperform current approaches in many situations, e.g. image or speech processing. One of the most important deep architecture is represented by the Convolutional Neural Network (CNN). The purpose of this paper is to provide practical recommendations in the deployment and development of the CNN based applications. They refer to the hardware as well as software available solutions and go beyond by providing guidance in choosing the appropriate hyper-parameters (structure, training algorithm, learning rate, regularization techniques, etc.). The experimental results are reported using the CIFAR-10 dataset.