Differentiable Topology Layers for Neural Networks

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2383 papers584 conceptsfrozen 2026-10-04

Standard deep learning excels at identifying local patterns—the edges in an image or the neighbors in a graph. However, these models often struggle with global structure: the presence of a hole in a manifold, the connectivity of a protein backbone, or the "loopiness" of an attention map. While Topological Data Analysis (TDA) provides tools like persistent homology to quantify these shapes, these tools were historically static, discrete, and incompatible with gradient-based optimization.

A growing body of research has bridged this gap by developing differentiable topology layers. These layers allow neural networks to not only "see" the global shape of data but to actively optimize it during training.

The Challenge of Topological Multisets

The primary output of TDA is the persistence diagram: a multiset of points $(b, d)$ representing the "birth" and "death" of topological features (like holes or components) as a filtration parameter changes.

As Mathieu Carrière and Frédéric Chazal point out in A General Neural Network Architecture for Persistence Diagrams and Graph Classification, these diagrams are difficult to use in deep learning because they have varying cardinalities. You cannot simply feed a list of 10 holes from one image and 50 holes from another into a fixed-size dense layer.

Early approaches, such as those by Zixuan Cang in A topological approach for protein classification, solved this by "vectorizing" the diagrams into fixed-length summaries like persistence landscapes or images. However, these were static preprocessing steps. The model could use the features, but it couldn't learn which topological features were most important for the task.

Making Topology Differentiable

The breakthrough came from realizing that the birth and death of a topological feature are determined by specific critical values in the input data—such as the pixel intensity at a certain coordinate or the weight of a specific edge in a graph.

In A Topology Layer for Machine Learning, Rickard Brüel Gabrielsson demonstrated that persistent homology is differentiable. By tracking which input parameters define the birth and death points, gradients can flow from a topological loss function back to the raw input. This allows for "topological regularization," where a network can be penalized if its output doesn't have the correct number of connected components or holes.

Similarly, Christoph D. Hofer and Roland Kwitt introduced a learnable projection in Deep Learning with Topological Signatures. Instead of using a fixed mapping, they treated the vectorization process as a neural layer with learnable parameters (like Gaussian structure elements). This allows the network to adaptively focus on specific regions of the persistence diagram that are most relevant to the target task.

Learning the Filtration

In graph learning, the "shape" of a graph depends on the function used to color its nodes (the filtration). Christoph D. Hofer and Bastian Rieck extended this in Graph Filtration Learning by making the filtration function itself learnable.

Instead of relying on hand-crafted properties like degree or centrality, the model learns a function that assigns values to nodes such that the resulting topological signature is optimal for classification. This turns persistent homology into a "readout" operation that captures global structural features that standard message-passing Graph Neural Networks (GNNs) might miss.

Applications in Structure and Geometry

The intuition behind these layers is most visible in fields where "shape" is synonymous with "function":

  • Protein Biology: Aditi S. Krishnapriyan and colleagues developed PersGNN, which uses learnable topological layers to predict protein function. By optimizing the representation of 3D voids and loops, the model captures the geometric "pockets" where drugs might bind.
  • Latent Space Regularization: In Connectivity-Optimized Representation Learning via Persistent Homology, Christoph D. Hofer used a differentiable connectivity loss to force the latent space of an autoencoder to follow a specific topological structure, significantly improving performance in one-class learning and anomaly detection.
  • Image Segmentation: Dimitris Samaras and Chao Chen showed in Learning Probabilistic Topological Representations Using Discrete Morse Theory that by making the sampling of persistence thresholds differentiable, models can learn to produce segmentations (like blood vessel networks) that are topologically correct, avoiding the "broken" connections common in pixel-wise classifiers.

By treating topology as a first-class, differentiable citizen in the neural network architecture, this research direction moves beyond simple pattern matching toward a deeper, structural understanding of data.

Papers in this version

Showing 60 of 383 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 234 papers
  2. Version 1 · 2026-10-04 · 383 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_topological_vectorization_4b59ad8f,
  title = {topological vectorization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/topological-vectorization/rdr_4b59ad8ff322488dbff05d0a2557370d},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_differentiable_topology_layers_for_neura_4b59ad8f_v1,
  title = {Differentiable Topology Layers for Neural Networks},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/differentiable-topology-layers-for-neural-networks/rdr_4b59ad8ff322488dbff05d0a2557370d/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}