Compressed Particle-Based Federated Bayesian Learning and Unlearning

Conventional frequentist federated learning (FL) schemes are known to yield overconfident decisions. Bayesian FL addresses this issue by allowing agents process and exchange uncertainty information encoded in distributions over the model parameters. However, comes at cost of a larger per-iteration communication overhead. This letter investigates whether can still provide advantages terms calibration when constraining bandwidth. We present compressed particle-based protocols for “unlearning” that apply quantization sparsification across multiple particles. The experimental results confirm benefits robust bandwidth constraints.