Reproduce “Adaptive Bayesian Online Learning via Expert Aggregation” 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
Jungbin Jun, Ilsang Ohn
A framework for adaptive Bayesian online learning integrates Bayesian updates with online expert aggregation, allowing dynamic adjustment of inferential choices like learning rates and prior distributions. The approach demonstrates enhanced robustness in nonstationary data streams and achieves minimax-optimal adaptation for unknown smoothness in Gaussian Process regression.
Key insights
- The choice of expert scoring mechanism, mean-loss or annealed-loss, determines the regret bound, with annealed-loss providing a logarithmically faster rate for cumulative predictive loss.
- Selecting the appropriate scoring rule should align with the specific inferential objective, such as optimizing posterior mean functionals or tracking full predictive distributions.
- Expert aggregation offers a modular approach to achieve sophisticated adaptive properties, like minimax-optimal rates for unknown smoothness in Gaussian Processes, without complex hierarchical Bayesian model construction.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.