PDA: A unified framework for lossless, one-shot, federated learning algorithms

Multi-site studies are now central to biomedical research, but inference across sites remains challenging due to regulatory limits on sharing individual-level data, heterogeneous data distributions, sparse events in small centers, and the logistical burden of multi-round communication.

In this talk, I introduce MOSAiC (Multi-site One-Shot Aggregation of Compressed Risk Functions), a unified framework for modern distributed research networks. Notably, MOSAiC reframes distributed learning as a mathematical problem of compressing and aggregating local risk functions, leveraging recent advances in tensor networks: state-of-the-art tools for high-dimensional function approximation in scientific computing.

Our MOSAiC enjoys four desirable properties that have never been achieved by any of the existing federated algorithms (except linear regressions): one-shot communication, lossless recovery of pooled-data estimates, inclusiveness of all sites irrespective of size or event rarity, and analytic submodel exploitability without re-querying partners. I will illustrate MOSAiC's validity and efficiency through applications in drug relabeling, drug repurposing, and post-market safety surveillance.