Reproduce “BTFL: A Bayesian-based Test-Time Generalization Method for Internal and External Data Distributions in Federated learning” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yu Zhou, Bingyan Liu

A Bayesian-based method, BTFL, addresses the challenge of balancing personalization and generalization for federated learning models during online inference by adapting to both internal and external data distributions at test time. It achieved average accuracy improvements of up to 3.15% over baselines and processed samples 4X-12X faster, offering computational efficiency for edge devices.

Key insights

  • A Bayesian framework provides a theoretically grounded approach to dynamically balance personalized and generalized model predictions for individual test samples.
  • Leveraging discretized likelihood estimators (DLEs) and entropy-based heuristics enables robust Bayesian updates for characterizing data distribution shifts at test time.
  • An optimization-free calculation of the interpolation coefficient, based on the posterior expectation of a mixing probability, achieves high computational efficiency during inference.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.