In plain words: It pulls together a mathematical framework that breaks down what each layer of a picture-recognizing network actually computes. Instead of treating these networks as black boxes judged only by accuracy, it shows step by step why their filtering and pooling design works so well.

Abstract

Convoulutional Neural Networks (CNNs) exhibit extraordinary performance on a variety of machine learning tasks. However, their mathematical properties and behavior are quite poorly understood. There is some work, in the form of a framework, for analyzing the operations that they perform. The goal of this project is to present key results from this theory, and provide intuition for why CNNs work.

Jayanth Koushik
arXiv:1605.09081 · stat.OT · submitted May 30, 2016
abstract · pdf · html · Statistical Machine Learning Course Project at Carnegie Mellon University