Topologically Preserving Neural Representations via Homeomorphic Mappings

This direction is merged into another direction: topology-preserving deformation. Its last version stays readable and citable here.

Version 1 of 112 papers21 conceptsfrozen 2026-10-04

Standard deep learning architectures are often described as "black boxes" that warp and twist input data into a form suitable for classification. While effective, this process frequently results in topological distortion: the network may "tear" the data space or "fold" distinct clusters onto one another. When these geometric relationships are destroyed, models lose their ability to generalize across domains or memorize specific points without affecting the broader input space.

A growing body of research proposes a more disciplined approach: constraining neural transformations to be homeomorphisms. By ensuring that mappings are bijective (one-to-one) and continuous in both directions, these models preserve the underlying "shape" of the data throughout the learning process.

The Problem of Topological Distortion

In a typical feed-forward network, layers often compress data into lower-dimensional spaces. This compression can be destructive. As noted by Cody Hyndman in The NEU Meta-Algorithm for Geometric Learning with Applications in Finance, standard architectures lack constraints to prevent irreversible distortions like folding or tearing.

If a model "tears" the space, two points that were originally close together might end up in completely different regions of the representation space. Conversely, if it "folds" the space, the model might collapse distinct class clusters into the same coordinate, making them inseparable. This lack of topological integrity prevents the model from maintaining the geometric relationships necessary for robust generalization.

Homeomorphisms as Structure Preservers

To solve this, researchers use homeomorphic mappings. A homeomorphism is a function that acts like a continuous deformation—think of stretching a rubber sheet without cutting it or gluing parts together.

The primary mechanism for achieving this is the Invertible Neural Network (INN). Because INNs are mathematically guaranteed to be bijective, they provide a one-to-one correspondence between the input and the latent space. As Lihua Zhou and colleagues demonstrate in Homeomorphism Alignment for Unsupervised Domain Adaptation, this bijectivity ensures that the topological structure and discriminative clusters of the data are preserved. If two points belong to different categories in the source domain, a homeomorphic mapping ensures they remain distinct and coherent in the target domain.

Upgrading Models with Non-Euclidean Features

One powerful application of this concept is Non-Euclidean Upgrading (NEU). In NEU: A Meta-Algorithm for Universal UAP-Invariant Feature Representation, Cody Hyndman introduces "Reconfiguration Networks."

Instead of standard weights, these networks utilize matrix exponentials of skew-symmetric matrices to construct higher-dimensional rotation matrices. This design ensures the feature map is a homeomorphism. By "upgrading" a simple linear model with this specialized feature map, the model becomes a universal approximator that can memorize specific data points without distorting the rest of the input space. This allows for high-precision learning that respects the original data's global geometry.

Bridging Domains Without Information Loss

Topological preservation is particularly vital in Unsupervised Domain Adaptation (UDA), where a model must transfer knowledge from a labeled source (e.g., synthetic car images) to an unlabeled target (e.g., real drone footage).

Traditional UDA methods often force both domains into a "common space," which can mangle their internal structures. Recent work offers alternatives:

Modeling the Dynamics of Inference

Beyond improving performance, these concepts help us understand how neural networks function. In Koopman Autoencoders Learn Neural Representation Dynamics, Nishant Suresh Aswani and Saif Jabari treat the layer-by-layer transformation of data as a dynamical system.

They argue that for a surrogate model to accurately replicate a deep network's behavior, it must preserve the geometric and topological structures (like clusters or "holes" in the data manifold) during the mapping process. When this topological fidelity is maintained, the surrogate can effectively "steer" the network's dynamics, enabling tasks like targeted class unlearning without retraining the entire model.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_topologically_preserving_neural_represen_3c2edb79,
  title = {Topologically Preserving Neural Representations via Homeomorphic Mappings},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/topologically-preserving-neural-representations-via-homeomorphic-mappings/rdr_3c2edb79a6854c658360406759fc57f2},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_topologically_preserving_neural_represen_3c2edb79_v1,
  title = {Topologically Preserving Neural Representations via Homeomorphic Mappings},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/topologically-preserving-neural-representations-via-homeomorphic-mappings/rdr_3c2edb79a6854c658360406759fc57f2/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}