New research dissociates function and representation in deep linear neural networks

Researchers have analyzed deep linear neural networks to understand the relationship between their representations and functions. They discovered that functional and representational similarity are not always aligned, meaning networks can have similar representations without performing similar tasks, and vice versa. The study also found that robustness to parameter noise, rather than robustness to input noise or generalization error, constrains representations to be task-specific. These findings suggest that representational alignment offers computational advantages beyond just functional alignment, impacting how we interpret and compare connectionist systems. AI

IMPACT Provides theoretical insights into the relationship between neural network representations and functions, potentially guiding future model interpretability and design.

RANK_REASON The cluster contains a single academic paper published on arXiv, detailing theoretical research into neural networks. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• alphaXiv
• arXiv
• CatalyzeX Code Finder for Papers
• connectionism
• CORE Recommender
• cs.LG
• DagsHub
• Gotit.pub
• Hugging Face
• IArxiv Recommender
• Multivariate pattern analysis of fMRI: The early beginnings
• ScienceCast

• paper
• other

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