Identifying classifier-relevant regions in images through weightless learning

Abstract

The need for eXplainable Artificial Intelligence becomes apparent as deep learning models grow in popularity and Artificial Intelligence is used in more and more areas. Many techniques have been proposed thus far to produce humanlegible explanations to the decision processes of classifiers, each solving a small piece of this enormous puzzle. One part of this, are the visual explanations, which seeks to produce images which can highlight what is the relevant content to the classifier, and clue the user in as to whether the model makes its decision on a sound basis, or at random, or even on a mistaken premise. Thus, solutions such as LIME find ways to generate theses explanations across different learning models, providing a versatile tool to better understand classification models. Although said solutions usually attempt to be as model agnostic as possible, the natural caveat is that they are better suited for some classes of problems and classifiers than others. Therefore, a number of different explainable models are needed in order to cover the vast space of possible models to explain. We introduce one such model, the Fuzzy Regression WiSARD Interpreter (FRWI), to attempt to produce higher quality explanations from WiSARD based models. Furthermore, as we need an objective, quantifiable way of gauging how different models compare, we also introduce our own Interpretation Capacity Score (ICS), a measurement process to judge the explanations produced. Under this metric as well as subjective, qualitative tests, this new FRWI approach had promising results, which could beat LIME in the tested scenarios and provide comprehensible explanations.