🔬 condensed matter

Shortcut learning in geometric knot classification

This paper investigates how machine learning models often rely on non-topological shortcuts rather than true topological features to classify geometric knots, and addresses this by providing a rigorous, publicly available dataset and code designed to eliminate such biases and foster the development of robust ML solutions for knot classification.

Original authors: Djordje Mihajlovic, Davide Michieletto

Published 2026-02-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Djordje Mihajlovic, Davide Michieletto

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

Imagine you are trying to teach a computer how to tell the difference between a perfectly smooth, untangled loop of string (an "unknot") and a tangled, knotted loop (like a trefoil knot). This is a classic problem in mathematics called "knot theory."

Recently, scientists tried using Artificial Intelligence (AI) to solve this. They fed the AI thousands of pictures of knots and watched it learn. The AI seemed to get it right almost 100% of the time! Everyone was excited.

But this new paper says: "Wait a minute. The AI isn't actually learning what a knot is. It's cheating."

Here is the story of how they discovered the cheat code, using some simple analogies.

1. The "Cow on a Hill" Problem

In the world of AI, there is a famous example of "shortcut learning." Imagine you train an AI to recognize cows. You show it thousands of photos of cows. But, by accident, every single photo of a cow in your training set has a green grassy hill in the background.

The AI doesn't learn to look for the cow's shape, ears, or tail. Instead, it learns a shortcut: "If I see a green hill, it's a cow!"

If you show it a cow in a barn, it fails.
If you show it a green hill with a sheep on it, it thinks it's a cow.

The AI learned a pattern that usually works, but it didn't learn the truth.

2. The Knot Cheating

The authors of this paper found that the AI models used to classify knots were doing the exact same thing.

The Setup: The researchers used data generated by Molecular Dynamics (MD) simulations. Think of this as a computer simulation of a physical string made of beads. To keep the string from breaking or passing through itself, the simulation uses "physics rules" (like energy and temperature).
The Cheat: Because of these physics rules, the "untangled" strings (unknots) in the simulation tended to be small and tight, while the "tangled" strings (knots) tended to be larger and looser.
The Result: The AI didn't learn the complex math of how a knot is tied. It just learned: "If the string is big, it's a knot. If it's small, it's not."

It was like the AI was measuring the size of the ball of yarn rather than looking at the knot inside it.

3. The "GEOKNOT" Trap

To prove the AI was cheating, the authors built a new, stricter training ground called GEOKNOT.

The Old Way (MD): Like a child playing with a ball of yarn in a small, crowded room. The yarn gets stuck in certain shapes because there isn't enough space to move freely.
The New Way (GEOKNOT): Like letting that same child play with the yarn in a giant, empty warehouse. The yarn can stretch, twist, and curl in any shape, big or small, without being forced by physics.

In this new "warehouse" dataset:

The "untangled" strings could be huge.
The "tangled" strings could be tiny.
The old "size" shortcut no longer worked.

The Result: When the researchers tested the "smart" AI models on this new data, they failed miserably. Their accuracy dropped from 99% down to about 50% (which is basically guessing). This proved the AI never actually learned what a knot was; it just memorized the size of the strings in the old training set.

4. The "Magic Wand" Test

To be absolutely sure, the authors did one more test. They took a "fake" knot that the AI thought was a real knot (because it was big and twisty) and used a mathematical "magic wand" to smooth it out.

They smoothed the string until it became a perfect, simple circle (an unknot).
The Catch: The AI's prediction changed during the smoothing process. As soon as the string got smaller or less twisty, the AI said, "Oh, now it's an unknot!"
The Truth: A real mathematician knows that if you smooth a knot without cutting it, it stays a knot (or stays an unknot). The AI's answer should never have changed. The fact that it did changed proves the AI was looking at the shape (geometry), not the structure (topology).

5. Why Does This Matter?

You might ask, "So what? The AI got 99% right on the old tests."

The authors argue that this is dangerous for science.

In Physics and Biology: Knots appear in DNA, proteins, and polymers. If we use AI to study them, we need to know if the AI is understanding the structure of the DNA or just guessing based on how "loose" the simulation looks.
The Future: The paper provides a new toolkit (the GEOKNOT dataset) to train AI models that actually learn the math of knots, rather than just memorizing the size of the string.

The Bottom Line

The AI models were like students who memorized the answers to a specific test but didn't understand the subject. When the teachers changed the test slightly (by using the GEOKNOT dataset), the students failed.

This paper is a wake-up call: Just because an AI gets a high score, doesn't mean it understands the concept. We need to make sure our AI is learning the rules of the game, not just the tricks of the specific players.

1. Problem Statement

The classification of knot topology (determining if two closed curves are equivalent under ambient isotopy) is a fundamental problem in low-dimensional topology with applications in physics and biology (e.g., protein folding, DNA). While Machine Learning (ML) has shown promise in classifying knots, recent high-accuracy results (>99%) have raised concerns about shortcut learning.

The Core Issue: ML models trained on data generated by Molecular Dynamics (MD) simulations may not be learning true topological invariants. Instead, they may be exploiting non-topological geometric features (such as the physical size or specific writhe distributions) that are inadvertently correlated with knot types due to the energy constraints and sampling limitations of the MD simulations. This leads to models that fail when presented with geometrically complex but topologically identical embeddings (Out-of-Distribution or OOD failure).

2. Methodology

A. Shortcut Probe Algorithm

To diagnose whether models are learning topology or geometric shortcuts, the authors developed a diagnostic tool based on Mutual Information (MI).

Concept: They compute the MI between various geometric functionals ( $\phi_j$ ) and the knot labels ( $y$ ).
Functionals: These include total space writhe ( $\Omega^+$ ), total curvature ( $\kappa^+$ ), sum of pairwise distances ( $\Sigma^+$ ), and long-range entanglement metrics.
Logic: If a geometric feature has high MI with the label in a specific dataset, it suggests the feature is a "shortcut" the model could exploit. A true topological classifier should rely on features invariant under ambient isotopy, not transient geometric properties.

B. Dataset Generation

The study compares three distinct datasets to test generalization:

MD Low Temperature ( $T=1$ ): Generated using LAMMPS with standard bead-spring models. Represents the regime used in most existing literature.
MD High Temperature ( $T=10$ ): Same MD setup but with higher thermal energy to broaden the conformational space.
GEOKNOT: A novel, open-source Monte Carlo sampler developed by the authors.
- Mechanism: Uses a combination of BFACF (local, topology-preserving moves) and Pivot (non-local, topology-altering moves) algorithms on a lattice.
- Biasing: Unlike MD, GEOKNOT explicitly biases sampling to explore a broad, flattened distribution of geometric features (e.g., forcing high/low writhe and crossing numbers) regardless of energy constraints.
- Validation: Topology is rigorously checked at every step using Alexander polynomials and Vassiliev invariants.

C. Model Architecture

Model: Simple Feed-Forward Neural Networks (FFNN) with 4 layers (320 neurons).
Inputs: Tested on two representations:
1. Raw 3D coordinates of polygonal segments.
2. Writhe matrices (discretized space writhe density).
Training: Models were trained on MD datasets and tested on both MD and GEOKNOT datasets to measure generalization.

3. Key Results

A. Failure to Generalize to GEOKNOT

Models trained on MD datasets achieved >99% accuracy on MD test sets.
However, when tested on the GEOKNOT dataset (which contains geometrically diverse embeddings of the same knots), accuracy dropped to ~50% (random chance) for coordinate-based models and ~67% for writhe-matrix models.
Conclusion: The high accuracy on MD data was not due to learning topology, but rather learning the specific geometric constraints (shortcuts) inherent to MD sampling.

B. Identification of Shortcuts

Mutual Information Analysis: In MD datasets, geometric functionals like total space writhe ( $\Omega^+$ ) and pairwise distance sums ( $\Sigma^+$ ) showed high mutual information with knot labels. In GEOKNOT, these correlations vanished (MI $\approx$ 0).
Shortcut Index ( $\tau$ ): The authors defined $\tau = m_a / m$ $τ = m_{a} / m$ , where $m_a$ $m_{a}$ is accuracy using only shortcut features and $m$ $m$ is accuracy using full data.
- For MD Low Temp, $\tau \approx 1.0$ , indicating the model relies almost entirely on geometric shortcuts.
- For GEOKNOT, the model failed to classify even when using the "shortcut" features, confirming the lack of correlation in that dataset.

C. Sensitivity to Ambient Isotopy

Ablation Study: The authors took GEOKNOT unknots (01) that were misclassified as trefoils (31) by MD-trained models and minimized them using KnotPlot to remove non-topological crossings.
Result: As the total space writhe decreased during minimization, the model's prediction probability for "trefoil" dropped, and it correctly classified the knot as an "unknot."
Implication: The model's decision boundary is sensitive to geometric deformations that do not change topology, proving it does not understand ambient isotopy.

D. Validation of Published Models

The authors tested existing published models (Sleiman et al., Braghetto et al., Zhang et al.) on the GEOKNOT dataset.
Result: All existing models failed significantly (accuracies ranging from 14% to 52%), confirming that shortcut learning is a systemic issue in current knot classification literature.

E. Potential for True Topological Learning

The authors demonstrated that Vassiliev invariants (specifically the second invariant, $V_2$ ) can be approximated directly from writhe matrices using algebraic manipulations derived from Chern-Simons theory.
Applying this rule-based calculation to GEOKNOT data yielded ~98% accuracy, proving that the geometric data contains the necessary topological information, but standard ML models fail to extract the complex algebraic structure required to learn it.

4. Key Contributions

Diagnosis of Shortcut Learning: Provided rigorous evidence that current ML success in knot classification is largely an artifact of dataset bias (MD energy constraints) rather than genuine topological understanding.
GEOKNOT Dataset: Introduced a publicly available, open-source sampler that generates knot embeddings with controlled, broad geometric distributions, serving as a robust benchmark for future ML models.
Diagnostic Framework: Developed a "shortcut probe" based on mutual information to evaluate datasets before training, helping researchers identify potential biases.
Theoretical Insight: Showed that while writhe matrices contain sufficient information for topological classification (via Vassiliev invariants), current neural network architectures are insufficient to learn the necessary non-linear algebraic mappings.

5. Significance

This paper fundamentally shifts the paradigm for applying ML to topology and physics:

For ML in Physics: It warns against trusting high accuracy on physically constrained datasets without OOD testing. It highlights that models may learn "physics" (energy constraints) rather than "topology" (invariants).
For Knot Theory: It clarifies that geometric features (like writhe) are not topological invariants but can be highly correlated in specific physical regimes.
Future Directions: The work suggests that to truly learn topology, ML models must be trained on geometrically unbiased data (like GEOKNOT) and likely require architectures capable of discovering complex algebraic structures (like Feynman diagram contractions) rather than simple pattern matching.

The authors conclude that while ML is a powerful tool, current models are "shortcut learners" in this domain, and future efforts must focus on generating unbiased datasets and developing models capable of learning true topological invariants.