Transparent Machine Learning: Theory and Computation

Abstract

Modern machine learning is driven primarily by black-box models, which provide superior performance but offer limited transparency into how predictions are made. For applications where it is important to understand how models make decisions, and to assist in model debugging and data-driven knowledge discovery, we require tools that can answer questions about what influences a model’s behavior. This is the goal of explainable machine learning (XML), a subfield that develops tools to understand complex models from various perspectives, including feature importance, concept attribution and data valuation. This dissertation presents several contributions to the field of XML, with the main ideas organized into three parts: (i) a framework that enables a unified analysis of many current methods, including their links with information theory and model robustness; (ii) a suite of techniques to accelerate the computation of Shapley values, which are the basis of several popular algorithms; and (iii) a range of methods for performing feature selection with deep learning models, e.g., in unsupervised and adaptive settings. Many of these ideas are motivated by applications in computational biology and medicine, but they also represent fundamental tools and perspectives that are useful across a variety of domains.