Topics in the theory of interactive learning with explanations

The purpose of this research program is to advance the theory of interactive machine learning with explanations. The theory of machine learning has classically studied algorithms that learn a machine learning model by consuming data provided by an external data source. The data in these classical frameworks is limited to examples and their categorizations, with no additional information. For instance, if the goal is to learn which products a user likes, in the classical model the input to the learning algorithm consists only of examples of products and whether the user likes each of them. However, this ignores the possibility of obtaining a much higher-level feedback from the user, that can help predict their likes and dislikes faster and more accurately: users can explain why they like a certain product or dislike another. The goal of this research program is to provide a deep understanding of the ways in which explanation feedback can improve machine learning outcomes. This goal is motivated by the current status of Artificial Intelligence (AI). While AI has been immensely successful, it faces significant issues, such as requiring massive amounts of data, computation and energy, and being hard to audit for accountability. Using explanations to guide the learning process can create models that are easy to interpret by humans, while reducing the need for data, computation and energy. The long-term vision of this research program is to provide a comprehensive analytical framework for deriving explanation-based learning algorithms and analyzing their guarantees, and for evaluating the efficacy and usefulness of explanation-based interactive protocols under different types of learning scenarios. The short-term objectives include studying scenarios for learning with explanation feedback, designing algorithms and providing analysis that helps identify the capabilities of the algorithms and the benefits and disadvantages of each scenario. We will provide theoretical analyses of proposed methods, and also demonstrate their success by implementing them, running them on a wide range of data sets, and analyzing the results to understand their efficacy. Understanding how to incorporate explanations into the learning process is necessary to facilitate a better alignment of AI systems with the needs of Canadian society. This research will lead to breakthroughs for developing AI with a sustainable energy footprint, while achieving interpretability and accountability that are crucial for fairness and equity in the use of machine learning models. The ability to incorporate user explanations to personalize software systems and interfaces will lead to enhanced accessibility and equity in the use of technology for marginalized populations. The algorithmic and theoretical insights provided by this research program will further enrich the scientific community in the fields of artificial intelligence and machine learning.