Reproduce “Topology of Out-of-Distribution Examples in Deep Neural Networks” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Esha Datta, Johanna Hennig, Eva Domschot, Connor Mattes, Michael R. Smith
Researchers at Sandia National Laboratories explored the topological characteristics of out-of-distribution (OOD) data within the latent spaces of deep neural networks using topological data analysis. They found that OOD examples maintain distinct topological features, particularly in their H₀ homology, suggesting they do not undergo the same simplification process as in-distribution data.
Key insights
- Well-trained deep neural networks induce topological simplification on in-distribution data as it passes through layers, making it more separable.
- Out-of-distribution (OOD) examples do not undergo the same topological simplification process as in-distribution data within the latent space.
- Differences in topological features, particularly in H₀ (connected components) homology, are observable and statistically distinct between in-distribution and OOD data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.