Milestones and New Frontiers in Deep Learning

Abstract

Only very recently, deep learning models have been globally used as the basis for most of the on-going research on several fundamental computing areas, such as computer vision, information extraction, data generation and data understanding. The research in this field is still a black box for many people, particularly, regarding its more mathematical aspects. Users of high level packages often struggle to understand the reasoning behind the many building blocks of deep learning, such as convolutional layers, batch normalization and activation functions. In this tutorial, we seek to introduce the field of deep learning to a broad audience of students and professionals, walking through two important milestones of deep learning understanding: (1) the intuition behind the neural networks and batch gradient descent algorithm with back propagation, and (2) the convolutional network and its use as a feature extraction technique. These milestones lay a solid conceptual and mathematical foundation in which the many other deep learning concepts can be explored and built on. In addition, we will present new frontiers in deep learning techniques which can be used by the audience to gain a deeper understanding and to provide the key pointers to the areas that are currently being mostly sought in the literature.