Constrained Optimization and Machine Learning
Abstract
Decisions made by Machine Learning and Artificial Intelligence systems now, more than ever before, have direct impact upon the public. Everything from private companies to public services (e.g., healthcare) are becoming reliant upon decisions or information provided by automated predictive algorithms. Therefore, it is critical that engineers be able to create these systems so that they consider the real-world constraints connected to their decisions. Thus, emerging formulations of statistical learning models that require such risk-adaptive considerations give rise to constrained optimization problems. Specifically, as machine learning and data analytic pipelines become more complex, the individual learning algorithms that compose the pipeline are subject to increasingly complex requirements. These requirements can often be encoded naturally with constraints, where the learning algorithm is required to satisfy multiple properties at once. This type of thinking is common to risk management, where decisions must satisfy multiple criteria, and the formulation of such criteria is managed with great care so as to be robust and yield computationally attractive optimization problems. This mindset is naturally translatable to the machine learning setting, where learning algorithms are required to satisfy multiple criteria, while simultaneously optimizing an objective. In this work we explore new techniques for formulating and solving Constrained Empirical Risk Minimization (C-ERM) problems, where constraints are used to encapsulate important data-dependent performance specifications of the desired ML system. The technical tools of this work will be selected from new developments in constrained optimization, risk management, and Deep Learning (DL). In particular, new optimization methods and Neural Network (NN) architectures will be developed for solving important constrained ML problems. We will also work to develop new constrained modeling frameworks relevant to meta-learning, an emerging field related to small-data learning scenarios. Our work will strengthen the ability of engineers to train automated systems to account for real-world constraints and generate ML and AI systems that are robust, fair, and simply perform well in the real world. Approved for Public Release
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Apr 28, 2020
- Source ID
- N002442010005
Entities
People
- Harbir Antil
Organizations
- George Mason University
- United States Navy