Project Grant 2221871

The National Science Foundation Division of Computing and Communication Foundations awarded a $614,000 Project Grant to the University of California, Los Angeles to support research into efficient compression schemes for communication-constrained machine learning environments. Under the Computer and Information Science and Engineering program (CFDA 47.070), this three-year award will fund the development of novel techniques to compress data communicated between distributed learning agents, such as smartphones, drones, and vehicles, while preserving learning ability. Specifically, the university researchers will study fundamental bounds and efficient algorithms to minimize the number of bits communicated in compressing rewards, context vectors, state-action features, and models for multi-armed bandit, contextual bandit, and Markov decision problems. This work aims to advance the state of distributed online and active learning by improving communication efficiencies for low-capability devices performing collective learning over networks.