All projects Computer Vision

Persistent Homology

Measuring the shape of data across scales.

Study & implementation 2023 TopologyTDAFiltration
Persistent Homology diagram

Motivation

Topological data analysis asks a question orthogonal to most of machine learning: what is the shape of a dataset? Persistent homology answers by tracking topological features — connected components, loops, voids — as a scale parameter grows.

Method

A filtration builds a nested family of simplicial complexes over the data. Homology is computed at every scale, and features are summarized by persistence — how long each feature survives — visualized as a barcode or persistence diagram. Long-lived features are signal; short-lived ones are noise.

\[ \varnothing = K_0 \subseteq K_1 \subseteq \cdots \subseteq K_n = K,\qquad H_k(K_i) \to H_k(K_j) \]

Mathematical core

The stability of persistence diagrams makes topology a robust feature for learning. This project anchors my broader interest in geometric and topological structure — including using topological invariants to understand latent spaces and continuous-depth models.