Reproduce “A Topological "Reading" Lesson: Classification of MNIST using TDA” 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
Adélie Garin, Guillaume Tauzin
Researchers at EPFL's Laboratory for Topology and Neuroscience applied Topological Data Analysis (TDA) to classify MNIST handwritten digits, demonstrating that a significantly smaller set of TDA-derived features can achieve classification accuracy comparable to using raw pixel values. Their comprehensive pipeline systematically extracts and vectorizes topological features, achieving 96.3% accuracy on MNIST with 28 features, representing an approximate 5-fold dimensionality reduction compared to pixel-based methods.
Key insights
- Topological Data Analysis (TDA) can effectively extract intrinsic shape information from images, capturing structural features often missed by conventional pixel-based methods.
- A systematic comparison of various filtrations (e.g., Height, Radial, Density) and persistence diagram vectorizations (e.g., Persistent Entropy) reveals that certain topological features are highly discriminative for image classification tasks.
- TDA provides a powerful method for dimensionality reduction that yields features with geometric interpretations, enhancing model explainability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.