Mapping Machine Learning to Physics (ML2P)

The Defense Advanced Research Projects Agency (DARPA) is seeking proposals for the Mapping Machine Learning to Physics (ML2P) program. The goal of the program is to improve the power consumption and performance of machine learning models on existing hardware by preserving local energy semantics and tuning energy-performance objective functions. The program will focus on developing multi-objective optimization functions that balance power consumption with performance metrics, enabling energy-aware machine learning. Proposers will be required to submit an abstract, and if selected, a full proposal and participate in an oral presentation. The program consists of two phases, with a go/no-go decision at the end of Phase 1. The total budget for the program is anticipated to be at or below $5.9M, with multiple awards expected. Key aspects of the program include: * Developing multi-objective optimization functions for energy-aware machine learning * Mapping machine learning efficiency to physics using precise granular measurements * Enabling accurate predictions of power consumption and performance of machine learning models * Exploring the construction of salient semantic features for concisely capturing the needed energy semantics of machine learning * Developing software that generates energy-aware machine learning models optimized for both energy and performance * Publishing software documentation, algorithms, code, and tutorials as open-source via existing machine learning repository sites.