Conference on the Mathematical Theory of Deep Learning
Abstract
Recent advances in deep neural networks (DNNs), combined with standardized datasets [1-3], new hardware [4] and open, easily-accessible implementations [5,6], have made neural networks a powerful, versatile method used widely in computer science [7-9], engineering [10-14], neuroscience [15-22], psychology [23-26], and the physical sciences [27-30]. These practical advances, however, have far outpaced a formal understanding of these networks, their training [31,32], and their failure modes [33-39]. Failure modes are critical as DNNs are increasingly used in sensitive applications, e.g., healthcare [40-44] and autonomous vehicles [45-47]. The dearth of rigorous analysis also limits the use of DNNs in scientific inquiry, and hinders principled design of the next generation DNN methods. Recently, long-past-due theoretical results have begun to emerge from researchers in a variety of disciplines, wielding a diversity of frameworks [48]. The purpose of this conference is to give visibility to these results, to accelerate the emergence of others,and to revolutionize our understanding of these systems. This intensive two-day technical conference focuses on the state of the art in theoretical understanding of deep learning. The conference includes invited talks by world leading researchers across mathematics, physics, computer science, neuroscience, cognitive sciences and engineering, as well as contributed talks and a poster session. Topics of interest include: optimization, generalization, physics of deep learning, theory of recurrent neural networks and related topics.One of the most fundamental research questions today is how immensely complex and over-parameterized systems can yield state-of-the art performance. This question is both pivotal in modern-day Machine Learning and Artificial Intelligence, as well as a fundamental challenge to understanding likewise complex biological systems. Answers to this enigma will 1) unlock basic understanding both of how extremely complex systems, like the brain, can learn and interact sensibly with an equally complex world, 2) enable the principled design of even more powerful AI algorithms, and 3) improve the robustness, interpretability and performance in important application domains such as personalized medicine, autonomous vehicles, etc. To these ends, researchers across disciplines have been directing their efforts, leveraging methods from concentration inequalities [49] to statistical physics [50,51] to demystify such systems. DeepMath aims to provide a venue where researchers tackling this problem have a common meeting that is agnostic to the specificmethodologies, and instead focuses on this central question; Much like how neuroscience began with physicists, biologists, psychologists, engineers and mathematicians all interested in the same quandary. A similar space is now needed to build a complete theory ofover-parameterized learning systems that complements the emerging conferences that are discipline or application domain specific.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 13, 2025
- Source ID
- N000142512092
Entities
People
- René Vidal
Organizations
- Office of Naval Research
- United States Navy
- University of Pennsylvania