TL;DR of the paper, generated automatically.

This is an arXiv paper introducing HOLE, a method that uses persistent homology to analyze and interpret discriminative neural networks.

• How it works: HOLE extracts topological features from intermediate activations and presents them with cluster flow diagrams, blob graphs and heatmap dendrograms.
• What it's tested on: The authors evaluate it on a range of discriminative models, looking at representation quality, interpretability across layers, and robustness to input perturbations and model compression.
• Claimed findings: The abstract says topological analysis reveals patterns linked to class separation, feature disentanglement and model robustness, offered as a complementary perspective on understanding deep learning systems.
• Versions: First submitted 8 Dec 2025, with the current version (v4) revised 3 Nov 2026.