CAREER: Collaborative, Communal, and Continual Model Training for Democratizing Machine Learning

University Of North Carolina At Chapel Hill, Chapel Hill NC

Investigators

Abstract

This award is funded in whole or in part under the American Rescue Plan Act of 2021 (Public Law 117-2). Increasingly, much of the software used today is based on machine learning models. These models are programs whose behavior is learned from data. Creating these models can be computationally and economically expensive. Fortunately, many powerful models are shared openly by their creators. The most popular models are re-used millions of times and form the backbone of many important pieces of software. However, these models are seldom updated - they are often left frozen in their originally-released form. Furthermore, the cost of creating these models excludes most people from their development. In contrast, the paradigm of open-source software development provides a means for a distributed community to collaboratively build software. This kind of collaborative development is currently infeasible for machine learning models. The goal of this award is to develop the fundamental research to enable community-developed and continually-improved models. Moreover, in order to foster a diverse community of contributors, this award includes collaborative educational activities that include researchers from groups that are historically underrepresented in the development of machine learning models. Finally, the award contains an educational component that focuses on the development of "role-playing" courses, where students present the material to one another from different perspectives. A major challenge to community-developed and continually-improved models is that the de facto status of gradient descent as an optimization method for machine learning models prevents any kind of collaboration because any change to the model involves updating all of its parameters. This award considers two alternatives: The first focuses on methods for allowing sparse and low-rank updates that are cheap to communicate and store and the second considers model architectures where submodels are devoted to specific capabilities that can be independently updated. In both cases, it will be made possible to "merge" conflicting updates that were proposed by contributors performing training in parallel. Separately, the ability to propose updates to a model necessitates the ability to determine whether a given proposed update should be accepted. This award therefore includes a parallel focus on rapid adaptation and evaluation. Throughout the duration of the award, a software framework will be developed that enables true iterative version control of machine learning models. The research advances in the award will therefore see direct practical application. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.