Programming Massively Parallel Processors, 4th Edition

Chapter 16

Deep learning

With special contributions from Carl Pearson and Boris Ginsburg

Abstract

This chapter starts by introducing machine learning tasks. We then dive deeper into the classification task and introduce the perceptrons, a type of linear classifier that is foundational for understanding modern convolutional neural networks (CNN). We discussed how the forward inference and backward propagation training paths are implemented for both single-layer and multilayer perceptrons. On the basis of the conceptual and mathematical understanding of perceptrons, we present a basic convolutional neural network and the implementation of its major types of layers. We then present a CUDA kernel implementation of the convolutional layer, the most ...