Open Problems in Topological Machine Learning
Published:
You May Also Enjoy
TDA for Neuroscience: Topology of Brain Connectivity
less than 1 minute read
Published:
The brain is a network with rich topological structure. Clique topology (Reimann et al., 2017) used persistent homology to show that neocortical microcircuits form high-dimensional simplicial structures, far beyond random. PH-based connectivity fingerprints distinguish subjects, tasks, and disease states from fMRI. This post covers the key findings and open questions in topological neuroscience.
TDA for Biology and Genomics: Shape in the Life Sciences
less than 1 minute read
Published:
Biology is full of shape: protein folding, cell morphology, gene expression manifolds. TDA has found major applications in single-cell RNA-seq analysis (Mapper for trajectory inference), protein structure comparison (PH-based structural fingerprints), and cancer genomics (topological signatures of tumour heterogeneity). This post surveys the key results and tools.
TDA for Graphs: Persistent Homology Meets GNNs
less than 1 minute read
Published:
Persistent homology can be applied directly to graphs by defining filtrations on nodes or edges (e.g., by WL colours, degree, or learned scalars). The resulting persistence diagrams encode global graph topology — connectivity, cycles, cliques — beyond what standard 1-WL GNNs can detect. This post covers WL-filtrations, extended persistence on graphs, and hybrid GNN+PH architectures.
Topological Layers in Neural Networks: Differentiable TDA
less than 1 minute read
Published:
To train neural networks end-to-end with topological loss terms, we need to differentiate through the persistent homology computation. This post covers the gradient of persistence diagrams with respect to inputs (via sub-gradients through the reduction algorithm), differentiable vectorisations (PLLay, TopNet), and how to write a topological regulariser that encourages or discourages specific shape features during training.