Reproduce “Anchor-MoE: A Mean-Anchored Mixture of Experts For Probabilistic Regression” 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
Baozhuo Su, Zhengxian Qu
The Anchor–MoE architecture from Institut Polytechnique de Paris and Aerie Intelligent Technology Corp. introduces a two-stage Mixture of Experts for probabilistic regression, learning full predictive distributions with calibrated uncertainty. It achieved minimax-optimal L2 risk rates for the predictive mean and demonstrated competitive or superior performance in negative log-likelihood (NLL) against NGBoost on various UCI datasets.
Key insights
- Utilizing an initial "anchor mean" stabilizes the predictive base, allowing MDN experts to specialize in modeling heteroscedastic residuals, which improves overall predictive quality and variance stability.
- Combining a locality-aware metric-window kernel with a content-dependent soft router leads to effective expert specialization and load balancing, contributing to more accurate and robust predictive distributions.
- The model provides strong theoretical guarantees, including minimax-optimal L2 risk rates for the predictive mean and generalization bounds for CRPS, offering insights into statistical efficiency and guiding model sizing for practical applications.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.