New method enables sparse Bayesian neural networks with exact zero-weight sampling

Researchers have developed a novel method for creating sparse Bayesian neural networks (BNNs) using sticky piecewise deterministic Markov processes (PDMPs). This approach directly incorporates pruning within a Bayesian framework by employing spike-and-slab priors, which manage uncertainties related to both pruning and parameter estimation. The proposed PDMP-based sampling technique allows for asymptotically exact posterior sampling of BNNs, with the 'sticky' feature enabling weights to be set precisely to zero during the sampling process. This method has demonstrated that sparse BNNs can offer improved robustness against input noise compared to their dense counterparts. AI

IMPACT Introduces a novel sampling technique for sparse Bayesian neural networks, potentially improving robustness and efficiency.

RANK_REASON Academic paper detailing a new methodology for Bayesian Neural Networks. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv stat.ML →

• alphaXiv
• arXiv
• Bayesian Neural Networks
• CatalyzeX Code Finder for Papers
• DagsHub
• Filippo Pagani
• Gotit.pub
• Hugging Face
• Influence Flower
• Markov chain Monte Carlo
• Piecewise-Deterministic Markov Processes: A General Class of Non-Diffusion Stochastic Models
• ScienceCast

• paper
• model release
• infra

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