New research links parameter symmetries to representational geometry in neural networks

A new research paper published on arXiv explores the relationship between parameter symmetries and representational geometry in overparameterized nonlinear neural networks. The study demonstrates that a class of parameter symmetries can be broken down into three fundamental feature transformations: addition, duplication, and scaling. This analysis allows for a decomposition of representational geometry into essential and auxiliary components, clarifying how degeneracy can increase with overparameterization while maintaining fixed function. The research also suggests that specific implementation-level selection rules can resolve this degeneracy, leading to identifiable geometries that reflect the network's functional contributions. AI

IMPACT Provides a theoretical framework for understanding how representations in neural networks relate to their function, potentially guiding future model design and interpretation.

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

Read on arXiv cs.LG →

• artificial neural network
• arXiv
• feature transformations
• Hugging Face
• machine learning
• neuroscience
• Overparameterization
• parameter symmetries
• psychology
• Representational geometry: integrating cognition, computation, and the brain
• Representations

• paper
• model release

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