Project Grant 2231707

This three-year, $250,000 project grant from the National Science Foundation's Computer and Information Science and Engineering program will support research at Harvard University on approximate coded computing. The research aims to develop new theoretical frameworks and techniques to accelerate distributed machine learning computations, ensure accuracy in the presence of hardware errors and faults, and enable differentially private computations through controlled redundancy. Key products will include coding schemes for fault-tolerant approximate matrix multiplication and nonlinear computations, as well as differentially private computations. These techniques seek to combine ideas from information theory, coding theory, approximation theory, and differential privacy to establish fundamental bounds on tradeoffs between computation accuracy, data privacy, error tolerance, and redundancy overheads. Outcomes will be disseminated through publications, tutorials and curriculum to advance knowledge in this emerging area of coded computing for distributed machine learning.