Attribution methods for deep convolutional networks

Abstract

[en] In recent years, Deep Learning has shown great success across several areas. However, even, though it might provide remarkable accuracy for many tasks, its application in some fields faces a fundamental problem: its predictions are not interpretable. Attribution Methods offer a possible solution in regards to this problem. To do so, they resource to results in Game Theory in order to explain individual decisions made by Deep Learning algorithms. In this work, we will be focusing, specifically, on the application of Attribution Techniques to a subset of Deep Learning algorithms: Convolutional Neural Networks.