New SCROLL method optimizes Bayesian neural networks via Bethe free energy

Researchers have developed a new method for training Bayesian neural networks called SCROLL (Shared-Cavity fRee-rOuting Last-Layer). This approach optimizes the Bethe free energy rather than the typical evidence lower bound (ELBO), aiming for local consistency between network components. SCROLL is designed for single-pass training and can handle various likelihood functions, showing improved performance in predictive accuracy and calibration across several benchmark datasets. AI

IMPACT Introduces a novel training objective for Bayesian neural networks that may improve performance and efficiency.

RANK_REASON Academic paper detailing a new method for training Bayesian neural networks. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• Bayesian Neural Networks
• evidence lower bound
• Pavel Procházka
• SCROLL
• University of California, Irvine

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →