Deep learning Hessian spectra explained by hidden symmetry

A new paper published on arXiv explores the spectral properties of the Hessian matrix in deep learning models. Researchers have observed that eigenvalues in trained models tend to cluster, with a large group near zero and a few outliers. This study proposes that these patterns arise from a hidden, highly symmetric reference configuration. Modifications to the model architecture, data, or parameter metric break this symmetry, leading to the observed eigenvalue distribution. AI

IMPACT Provides a theoretical framework for understanding model behavior and potential avenues for architectural improvements.

RANK_REASON The cluster contains a research paper published on arXiv detailing theoretical findings about deep learning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• Convolutional Models for Landmine Identification with Ground Penetrating Radar
• deep learning
• Gauss-Newton matrix
• Graphical Models
• Hessian
• Relu Networks
• Transformer Models

• paper
• model release

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