🤖 machine learning

No Triangulation Without Representation: Generalization in Topological Deep Learning

This paper extends the MANTRA benchmark to diverse manifolds and reveals that while existing graph and higher-order models can saturate performance with proper representations, they fail to generalize beyond combinatorial structures, highlighting a critical gap in developing topology-aware inductive biases independent of scale.

Original authors: Johannes S. Schmidt, Martin Carrasco, Ernst Röell, Guy Wolf, Nello Blaser, Bastian Rieck

Published 2026-05-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Johannes S. Schmidt, Martin Carrasco, Ernst Röell, Guy Wolf, Nello Blaser, Bastian Rieck

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: The "Shape" versus the "Blueprint"

Imagine a ball of modeling clay. You can squeeze it, stretch it, or poke a hole in it, but as long as you don't tear it apart or glue it to something else, it still has the same "shape" (topologically speaking).

Now imagine you want to teach a computer to recognize this shape. The computer doesn't see the smooth clay ball; it only sees a blueprint made of glued-together triangles. This is called "triangulation."

The problem this paper addresses is that there are millions of different blueprints for exactly the same clay ball. You can draw the triangles large, small, or in a completely different pattern, and the underlying shape remains the same.

The authors ask: Do current AI models actually learn the shape, or do they just memorize the specific blueprint they have seen?

The Problem: The "Fake" Test

Previously, researchers tested these AI models using datasets consisting mainly of regular graphs (like social networks) that were artificially "lifted" to look like 3D shapes. This was like testing a pilot's ability to fly an airplane by having them drive a car painted to look like an airplane. The test was unfair.

The paper uses a dataset called MANTRA, which contains real, complex 3D shapes built from triangles. However, the authors identified two serious flaws in how this dataset was used:

Missing Labels: Many of the shapes had no names (e.g., "This is a sphere," "This is a donut").
The Wrong Test: The models were tested on how well they could distinguish between two different blueprints of the same shape, rather than between two different shapes.

The Solution: The "Remodeling" Experiment

To fix this, the authors did two things:

1. They filled in the missing labels.
They used mathematical rules (so-called Pachner moves) to generate thousands of new blueprints for each shape. Imagine taking a house blueprint and redrawing the walls, windows, and doors in different arrangements, but keeping the house exactly the same size and shape. This ensured the AI had enough examples to learn from.

2. They created a "stress test" (the refinement).
This is the most important part. They took the blueprints the AI was trained on and refined them.

The Analogy: Imagine teaching a student to recognize a cat by showing them a photo of a fluffy, orange cat. Then you test them by showing a photo of a hairless, black cat. If the student says, "That's not a cat because it looks different," they have failed. They memorized the photo, not the concept of a cat.
The Experiment: The authors took the AI's training data and added more and more triangles to the shapes (subdivision). The shape remained the same (it was still a sphere), but the blueprint became much more complex and looked very different from what the AI had seen during training.

The Shocking Results

The authors tested two types of AI models:

Standard Graph Models (GNNs): The "old-fashioned" way of looking at data.
Higher-Order Models (HOMP): The "new, sophisticated" way, specifically designed for complex shapes.

The Findings:

On the original data: Both model types performed wonderfully. They could easily distinguish a sphere from a donut.
On the "stress test" (refined data): Everyone failed.
- As soon as the blueprint changed slightly (even though the shape did not), the AI models became confused.
- Their performance dropped to the level of random guessing.
- Even the "sophisticated" models, which were supposedly supposed to understand topological structure, could not handle it.

The Conclusion:
The AI models did not learn the shape (topology). They only memorized the blueprint (combinatorics). They were like a student who memorized the answer key for a specific test but could not solve a new problem requiring the same logic.

The Lesson "No Triangulation Without Representation"

The title is a play on the famous phrase "No taxation without representation." Here it means: You cannot judge an AI's ability to understand shapes without considering how those shapes are represented to the AI.

The paper found the following:

The way you feed data to the AI (the "representation") is more important than the complexity of the AI model itself.
Sometimes a simple model with the right data view works better than a complex model with the wrong data view.
Crucially: No current model can generalize. All fail when the "blueprint" changes, even if the "shape" remains the same.

Summary for Everyday Life

Imagine these AI models as tourists visiting a city.

The old way: We showed them a map of the city and asked, "Is this Paris?" They memorized the map.
The new test: We gave them a different map of Paris (drawn by a different artist, with different street names) and asked, "Is this Paris?"
The result: The tourists said, "No, that's not Paris. The map looks different."

The authors say: We must build AI that understands the city (the topology), not just the map (the triangulation). Until we have done that, these models are not "topological" at all; they are just very good at recognizing patterns on specific blueprints.

Technical Summary: No Triangulation Without Representation: Generalization in Topological Deep Learning

Problem Statement
Despite growing interest in Topological Deep Learning (TDL) for higher-dimensional datasets (e.g., simplicial and cell complexes), there is no consensus on how to evaluate these models. Existing benchmarks, such as TopoBench, rely predominantly on graphs that have been "lifted" to higher-dimensional representations, thereby failing to capture the intrinsic properties of higher-dimensional spaces. Consequently, evaluations often reduce to graph-based tasks, undermining the credibility of TDL models and raising questions about the necessity of higher-dimensional representations.

A specific gap exists in the MANTRA dataset, which contains inherently higher-dimensional objects (combinatorial manifolds). However, the original dataset suffers from severe class imbalances (e.g., 80% of 2D triangulations are unlabeled, and 3D data is dominated by the 3-sphere) and lacks a rigorous evaluation protocol to distinguish between learning genuine topological structures and exploiting combinatorial artifacts. Furthermore, current models are not tested on their ability to generalize across different combinatorial representations (triangulations) of the same underlying topological space.

Methodology
The authors propose a comprehensive framework to extend the MANTRA dataset and establish a new evaluation protocol based on representation diversity and triangulation refinement.

Dataset Extension and Augmentation:
- Labeling: Using the classification theorem for closed surfaces and connected sums, the authors assign homeomorphism-type labels to previously unlabeled 2D manifolds.
- Balancing: They employ Pachner moves (local transformations that preserve homeomorphism) and connected sums to generate new triangulations. This creates three new dataset variants: 2D-unbalanced (9 classes), 2D-balanced (22 classes, 55,000 samples), and 3D-balanced (9 classes, 3,000–5,000 samples per class).
- Deduplication: To prevent data leakage, they apply heuristics including f-vector comparison, Weisfeiler-Lehman (WL) hashing of incidence graphs, and direct isomorphism checks.
Generalization Stress Test (Refinement):
To test whether models learn topological invariants rather than combinatorial specifics, the authors introduce subdivision operations that alter the combinatorial structure while preserving the underlying manifold (homeomorphism type).
- Stellar Subdivision: Iteratively adding vertices to maximal simplices.
- Barycentric Subdivision: Replacing simplices with sequences of strict inclusions.
- These operations generate "Out-of-Distribution" test sets where topology remains constant, but the number of vertices and edges increases, forcing models to generalize beyond the combinatorial distribution of the training data.
Experimental Setup:
- Models: A range of Graph Neural Networks (GCN, RG-GCN, Graphormer) and Higher-Order Message Passing (HOMP) models (SCCNN, CWN) are evaluated.
- Representations: Models are tested on the full simplicial complex ( $T$ ), the 1-skeleton ( $S$ ), the dual graph ( $D$ ), and the Hasse diagram ( $H$ ).
- Encodings: Various feature assignments are tested, including random uniform features, Node Degree, Random Walk Positional Encoding (RWPE), and Moment Curve (MC) embeddings.
- Protocol: Standard splits into training, validation, and test data are used, followed by evaluation on the refined subdivision datasets.

Key Results

Benchmark Saturation: Contrary to previous claims, standard GNNs (particularly Graphormer and RG-GCN) and HOMP models can saturate the classification benchmark on the original MANTRA data when provided with the correct representation (e.g., dual graph or Hasse diagram) and feature encoding (e.g., Moment Curve). In 3D, Graphormer on the dual graph with Moment Curve encoding achieves performance comparable to CWN.
Failure of Generalization: When evaluated on subdivided triangulations (refinement stress test), all models fail to generalize. Performance collapses drastically after the first few subdivision steps, often falling below random chance and even below a deterministic heuristic based on the Euler characteristic.
Combinatorial vs. Topological Learning: The results indicate that current models learn combinatorial artifacts specific to the training triangulations rather than topological invariants. Even models with access to full simplicial complexes (HOMP) do not outperform GNNs on the dual graph when using appropriate encodings, and none demonstrate the ability to recognize the same manifold under different triangulations.

Significance and Claims
The work argues that the field has moved too quickly toward developing complex architectures without first establishing models that can truly understand topological structures independently of scale or combinatorial representation.

Evaluation Reform: The authors propose that MANTRA should serve as a "smoke test" (analogous to MNIST) for TDL. Any model claiming to learn topological structures must first saturate this benchmark and, crucially, demonstrate generalization across refinements.
Importance of Representation: The study emphasizes that "No Triangulation Without Representation." The choice of representation (e.g., dual graph vs. complex) and feature encoding is often more critical to performance than the model architecture itself.
Research Gap: The findings reveal a fundamental gap: existing models do not learn topological structures but exploit combinatorial patterns. The authors call for a re-evaluation of model design to develop inductive biases that are genuinely topology-aware and capable of recognizing homeomorphism types independent of specific triangulation or mesh density.

The work provides the necessary framework (extended datasets and refinement protocols) to evaluate future models and aims to shift the community's focus from combinatorial expressivity to genuine topological generalization.