New method quantifies class disentanglement in neural networks

Researchers have developed a new method using the Intersection Euler Characteristic Profile to measure class disentanglement in neural network representations. This technique quantifies the separation of class-conditional point clouds and reveals that disentanglement is concentrated in the early stages of training and is primarily pairwise. The study found that augmentation is the only training choice that effectively separates classes, while weight decay compresses overlap without separation, and depth and width have no significant impact. The findings suggest that joint entanglement of class triples is generally lower than their strongest pairs, a pattern observed across various network architectures. AI

IMPACT Introduces a new metric for understanding neural network representations, potentially aiding in model interpretability and development.

RANK_REASON Academic paper on a novel method for analyzing neural network representations. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• cs.LG
• Hugging Face
• Intersection Euler Characteristic Profile
• Spearman rho

• paper
• other

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