A new paper explores the concept of "concept dimension" in neural representations, questioning common methods of measurement. Researchers demonstrate that iterative erasure counts, often used to quantify how many directions a neural network uses to encode information, can be manipulated by invertible reparameterizations. This suggests that these counts are not intrinsic properties of the concept but are dependent on the specific measurement procedure used. The study uses examples like V-JEPA2 features to illustrate how different optimization procedures can yield varying results even when the underlying prediction problem remains the same. AI
IMPACT Challenges current methods for understanding neural network representations, potentially influencing future research into model interpretability.
RANK_REASON The cluster contains an academic paper discussing a theoretical concept in neural network representations.
Read on Hugging Face Daily Papers →
- Hugging Face
- V-JEPA2
- Adam
- Gaussian function
- Iterative Erasure Count Is Not an Affine-Invariant Concept Dimension
- Moore--Penrose
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
