Educational Course “Introduction to Deep Learning Using the Intel neon Framework”

Abstract

The interest of researchers in deep learning is constantly increasing. Deep learning methods penetrate physics, biology and other problem areas. Often, scientists and engineers consider a deep model as a “black box”, allowing them to solve a problem on a specific dataset. However, there are typical schemes for the application of deep learning methods and standard recommendations for the pragmatic use of various deep models. In this paper, we present a new training course that provides a minimal theoretical basis and creates practical experience in applying deep learning for solving computer vision problems using the Intel neon Framework. The course studies the biological basis of artificial neurons, provides a classification of deep models, discusses constructing deep topologies and training methods of fully-connected, convolutional and recurrent neural networks, describes methods of unsupervised learning. The issues of efficient utilization of computational resources of modern clusters for training and testing deep models are also discussed. A distinctive feature of the course is the independence from a deep learning framework. All materials can be easily modified to use another framework. The paper formulates the main ideas of the course, describes its structure and provides the results of training over two years at the Lobachevsky State University of Nizhni Novgorod. Course materials (in Russian and English) are publicly available under Apache 2.0 license.