A new typed tensor language has been developed to facilitate shared-state federated computations, which combine client-local tensor operations with aggregated shared state. This language separates client-partitioned data from globally available values using two tensor sorts and tracks partitioned axes. It enables typed one-round programs to factor through shared tensors, with encoder components represented by aggregations or contractions and the decoder being shared-only. The system extends to multi-round programs with persistent shared state and supports server-side first-order and curvature-block updates for federated analytics and FedSGD. AI
IMPACT This new language could streamline the development and execution of federated learning models, potentially improving efficiency and scalability in distributed AI training.
RANK_REASON The cluster contains an academic paper detailing a new technical language for a specific type of computation. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →