Machine Learning Beyond Prediction - Extracting Insights and Guiding Actions

点击翻译按钮获取中文摘要

Owing to breakthroughs in supervised learning using deep neural networks, applications of machine learning (ML) have proliferated, spreading to countless industries and societal decisions. Amid this excitement, some fundamental obstacles are often ignored. While ML systems are typically trained to estimate conditional probabilities, in automated systems their primary purpose is to guide actions.  The discrepancy between predictions and decisions is just one among many mismatches between the supervised learning formalism and real-world goals. For example, often the purpose of training a model is not simply to make a prediction but rather to extract qualitative insights, such as causal inference, data clustering or outlier detection.However, when machine learning tools are applied to extract any of those insights, strong assumptions (such as the data being generated by some parameterized probability distribution) are used, often implicitly. Just the same, machine learning is applied far beyond the strict confines of those assumptions. On the other end, for deriving hardness results, most of the theoretical analysis of required resources (be it computational time or training sample sizes) refer to worst-case scenarios, therefore being overly pessimistic. Under that view, large neural networks seem doomed to fail. This project addresses three areas of discrepancy between ML formalism and such real world goals. 1)Interpretability of ML-based tools: Traditional measures of success, such as statistical accuracy and computational efficiency, do not suffice for human consequential applications, where society expects accountability and interpretability. I will analyze formal notions of interpretability and investigate how such notions effect prediction accuracy and render models amenable to monitoring. I will develop theoretical principles under which today's deep learning tools can be leveraged to confer insights beyond their predictive accuracy. 2) Guided selection of clustering algorithms: In spite of the major practical importance of unsupervised learning, current practical implementations of such tasks are very rudimentary. There exists no methodical guidance for clustering tool selection for a given clustering task. I shall address this crucial lacuna by developing methods to guide task appropriate choices of clustering paradigms. 3) Alternatives to worst-case analysis of ML tasks: Many optimization problems that arise in machine learning are NP hard. For example, the training of even small neural networks. Just the same, such problems are being handled routinely on real data for many applications. Experimental evidence suggests that this success relies on some "tameness" of practically arising data. We propose to address this theory-practice discrepancy by distilling structural properties of inputs that can be assumed to hold for naturally arising input data, while giving rise to efficient algorithms for solving hard problems on such inputs.