New SOAP-Bubbles method simplifies neural network weight uncertainty estimation

Researchers have developed a new method called SOAP-Bubbles to estimate structured weight uncertainty in neural networks, making it more efficient and easier to implement. This approach adapts the SOAP optimizer by running a variational method called IVON in the eigenspace of SOAP's preconditioner. The resulting technique, Eigenspace-VON (EVON), offers costs comparable to SOAP and has demonstrated superior results in language model pretraining compared to existing diagonal-covariance methods. AI

IMPACT Simplifies the estimation of expressive posterior distributions for deep learning models, potentially improving performance in tasks like language model pretraining.

RANK_REASON The cluster contains a research paper detailing a new method for neural networks. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• Adrian Robert Minut
• arXiv
• Eigenspace-VON
• Hugging Face
• Neural Networks
• SOAP optimizer

• paper
• infra

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