Harnessing collaborative intelligence in the Post-LLM world

Location: BBB 3725
Zoom: https://umich.zoom.us/j/99501392350
Meeting ID: 995 0139 2350
Passcode: aiseminar

Abstract: Collaborative data science methods help harness a collective intelligence from fragmented and siloed organizational assets or device eco-systems. The following are two major challenges that arise in unlocking collaborative intelligence: 1.) Maintaining a high resource-efficiency and 2.) enabling collaborative data acquisition . To that extent, I share my research on extremely resource-efficient LLM fine-tuning in the federated and siloed settings. I then provide methods for responsible acquisition of representations (embeddings) of data from multiple entities. This is as opposed to a lot of good work that has been previously done instead on model-sharing in the pre-LLM era. Another problem that I tackle along this path, are methods that post hoc convert a machine learning pipeline trained with informal (heuristic) assurances to those equipped with formal trustworthiness guarantees for representation release. I conclude with my research on the development of data markets, which are pivotal in enabling seamless data acquisition across siloed clients based on efficient federated optimization and optimal experimental design.

Bio: Praneeth Vepakomma is a Visiting Assistant Professor at Massachusetts Institute of Technology (MIT) in the Institute for Data, Systems and Society (IDSS) and is an Assistant Professor in MBZUAI. He has extensive industrial experience from his time at Meta, Apple, Amazon Web Services, Motorola Solutions, Corning and several startups. This includes his role as a Principal Staff Scientist at Motorola Solutions. He has won the Meta PhD research fellowship in Applied Statistics, the ADIA Lab Fellowship and an OpenDP Academic Fellowship at Harvard. He won the Financial Times Digital Innovation award for his research non-profit, two SERC Scholarships (for Social and Ethical Responsibilities of Computing) from MIT’s Schwarzman College of Computing. a Baidu best paper award for the FedML system, a best paper runner up award at FG, and several best paper awards at workshops in ML conferences. He has been a mentor for the Fatima Fellowship to support students in their career paths. He holds a PhD from Massachusetts Institute of Technology (MIT) and an MS in Mathematical and Applied Statistics from Rutgers University.