Year: 2020
Type: dissertation
Abstract: I propose a method which can visually explain the classification decision of deep neural networks(DNNs). Many methods have been proposed in machine learning and computer vision seeking to\nclarify the decision of machine learning black boxes, specifically DNNs. All of these methods try to\ngain insight into why the network “chose class A” as an answer. Humans search for explanations by\nasking two types of questions. The first question is, “Why did you choose this answer?” The second\nquestion asks, “Why did you not choose answer B over A?” The previously proposed methods are\nnot able to provide the latter directly or efficiently.\nI introduce a method capable of answering the second question both directly and efficiently.\nIn general, the proposed method generates explanations in the input space of any model capable of\nefficient evaluation and gradient evaluation. It neither requires any knowledge of the underlying\nclassifier nor uses heuristics in its explanation generation, and it is computationally fast to evaluate.\nI provide extensive experimental results on three different datasets, showing the robustness of my\napproach and its superiority for gaining insight into the inner representations of machine learning\nmodels. As an example, I demonstrate my method can detect and explain how a network trained to\nrecognize hair color actually detects eye color, whereas other methods cannot find this bias in the\ntrained classifier.\nI provide the details on how this framework can be applied to discrete data as well. I\nproposed a SoftEmbedding function to be employed in conjunction with the discrete embedding\nfunction. I show results on textual reviews demonstrating my method’s ability to find bias in learned\nclassifiers.\nFinally, I provide the results of a user study, measuring how much this feedback helps\nusers to improve their understanding of the network’s learned function in comparison with other\npossible method
Author: Seyedamirhossein Feghahati
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W3159809106