Program Summary
Machine learning has shown remarkable success across many application areas in recent years, leveraging advances in computing power and the availability of large sets of training data. It provides a tremendous opportunity to deploy data-driven systems in more complex and interactive tasks including personalized autonomy, agile robotics, self-driving vehicles, and smart cities. Despite dramatic progress, the machine learning community still lacks an understanding of the trade-offs and mathematical limitations of related technologies for a given domain, problem, or dataset. Lacking is a fundamental theoretical framework for understanding the relationships among data, tasks, resources, and measures of performance—elements that would allow understanding what tasks are best suited for machines and which tasks are not.
The DARPA Fundamental Limits of Learning (FunLoL) program seeks to develop methodologies to evaluate the capabilities of learning system designs and guide practical implementations based on a formal understanding of the boundaries of their performance. The program will investigate mathematical frameworks that provide quantifiable and generalizable measures of learning to enable the design of systems with verifiable properties. In addition, understanding of such properties would help characterize fundamental limits across existing and new machine learning paradigms and reveal how to value the trustworthiness of results in a wide variety of applications.