Symmetric Divergence and Normalized Similarity: A Unified Topological Framework for Representation Analysis
Researchers have developed a new toolkit for analyzing neural network representations using topological data analysis. This toolkit introduces Symmetric Representation Topology Divergence (SRTD) to address asymmetry issues in existing methods and provide more detailed structural diagnostics. Additionally, Normalized Topological Similarity (NTS) offers a standardized, scale-invariant metric for benchmarking across different scenarios, overcoming limitations of previous unbounded scores. AI
IMPACT Introduces novel metrics for evaluating and comparing neural network architectures, potentially improving model development and benchmarking.