Generalization, broadly construed, is the ability of machine learning methods to perform well in scenarios outside their training data.
Despite being a well-developed field with a rich history, contemporary phenomena – in particular those arising from deep learning, most specifically large image and language models – are well beyond our current mathematical toolkit and vocabulary. It is not merely that analyses are too loose to be effective; rather, the settings have drastically evolved from the standard statistical setting of similar training and testing data, as the following examples illuminate: self-driving cars may need to navigate unfamiliar and even private or inaccessible roads; image generation software is expected to provide compelling images from essentially arbitrary input strings, with human operators indeed enjoying breaking the training data mold; AlphaFold and related software make protein predictions for species unrelated to those in their training set; the list goes on, without even scratching the surface of large language models and algorithmic tasks.
The goal of this program is to bring together remote and local researchers, both in academia and industry, as well as across mathematical and applied disciplines, with common goals of (a) organizing and crystallizing gaps between the theory and practice of generalization, and (b) sparking collaboration towards a concerted effort to close these gaps.
Peter Bartlett (UC Berkeley), Daniel Hsu (Columbia University), Po-Ling Loh (Cambridge University), Toni Pitassi (Columbia University), Andrej Risteski (Carnegie Mellon University), Matus Telgarsky (New York University), Rich Zemel (Columbia University)
Long-Term Participants (tentative, including organizers):
Peter Bartlett (UC Berkeley), Daniel Hsu (Columbia University), Po-Ling Loh (Cambridge University), Toni Pitassi (Columbia University), Andrej Risteski (Carnegie Mellon University), Matus Telgarsky (New York University), Rich Zemel (Columbia University), Yusu Wang (UCSD), Varun Jog (Cambridge University), Misha Belkin (UCSD), Surbhi Goel (University of Pennsylvania), Johannes Schmidt-Hieber (University of Twente), Samory Kpotufe (Columbia University), Nati Srebro (TTI-C), Eunsol Choi (UT Austin), Fanny Yang (ETH Zurich), Vatsal Sharan (USC), Han Zhao (UIUC), Mahdi Soltanolkotabi (USC), Sivaraman Balakrishnan (Carnegie Mellon University), Ryan Tibshirani (UC Berkeley), Nika Haghtalab (UC Berkeley), Jennifer Listgarten (UC Berkeley)