Machine learning detects terminal singularities

Algebraic varieties are the geometric shapes defined by systems of polynomial equations; they are ubiquitous across mathematics and science. Amongst these algebraic varieties are Q-Fano varieties: positively curved shapes which have Q-factorial terminal singularities. Q-Fano varieties are of fundamental importance in geometry as they are"atomic pieces"of more complex shapes - the process of breaking a shape into simpler pieces in this sense is called the Minimal Model Programme. Despite their importance, the classification of Q-Fano varieties remains unknown. In this paper we demonstrate that machine learning can be used to understand this classification. We focus on 8-dimensional positively-curved algebraic varieties that have toric symmetry and Picard rank 2, and develop a neural network classifier that predicts with 95% accuracy whether or not such an algebraic variety is Q-Fano. We use this to give a first sketch of the landscape of Q-Fanos in dimension 8. How the neural network is able to detect Q-Fano varieties with such accuracy remains mysterious, and hints at some deep mathematical theory waiting to be uncovered. Furthermore, when visualised using the quantum period, an invariant that has played an important role in recent theoretical developments, we observe that the classification as revealed by ML appears to fall within a bounded region, and is stratified by the Fano index. This suggests that it may be possible to state and prove conjectures on completeness in the future. Inspired by the ML analysis, we formulate and prove a new global combinatorial criterion for a positively curved toric variety of Picard rank 2 to have terminal singularities. Together with the first sketch of the landscape of Q-Fanos in higher dimensions, this gives new evidence that machine learning can be an essential tool in developing mathematical conjectures and accelerating theoretical discovery.

Paper

Similar papers

Peer review

Reviewer VZKS7/10 · confidence 3/52023-06-21

Summary

The authors trained a feedforward neural network to detect $\mathbb{Q}$-Fano varieties in dimension 8, among toric varieties defined by weights in a bounded region. The resulting classifier achieves a high (95%) accuracy. The authors used this classifier to build a large dataset of (probably) $\mathbb{Q}$-Fano and visualize it. The high accuracy achieved also inspired the authors to derive a new (exact) algorithm to detect $\mathbb{Q}$-Fano varieties.

Strengths

The paper is very well written, it address a problem that is interesting and probably quite important (I am not an expert in algebraic geometry though). It has a very detailed experimental part, as well as a new theoretical result (which I haven't been able to check).

Weaknesses

It's not very clear whether, at the end of the day, the classifier built in this work is actually going to make an impactful contribution in the area. This weakness is outweighted by the strengths, in my opinion.

Questions

I have two provocative questions. 1. Was Proposition 5.3 (or its proof) inspired in any way by the neural network? The authors seem to imply it, but that does not seem very likely. To motivate in finding this theoretical result, it seems that the important thing was to strongly believe that a simple classifier exists, not that it actually exists. 2. Regarding the poor performance on out-of-distribution examples (Figure 4): Do you think this might mean that the trained neural network is not, in fact, good enough evidence that there is a simpler algorithm to detect $\mathbb{Q}$-Fano varieties? I also found one typo: - Line 293: "that" -> "than"

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

The paper does a good job of recognizing the weaknesses in the closing section.

Reviewer B8mu6/10 · confidence 4/52023-06-25

Summary

The authors study properties of algebraic geometric objects, four complex dimensional Fano toric varieties, using mathematical data science methods, i.e. statistical analysis of a large database of examples combined with mathematical domain knowledge. They find a fairly accurate neural network classifier to identify singular varieties and inspired by this propose a more efficient algorithmic test. They also study an invariant called the quantum period.

Strengths

Interesting example of mathematical data analysis for an open and actively studied problem in algebraic geometry. The new algorithm for singularity testing is interesting.

Weaknesses

The paper is very unsystematic statistically. Since the dataset consists of mathematical objects generated by a known sampling procedure, one could precisely characterize the ensemble being sampled in section 3 (with the bounds on weights given there), to help interpret the results and help compare with other (future) work. Or (probably better) change the sampling to get a simple to describe ensemble. This question is not even asked. It bears on the significance of the results and the reported poor performance on out of sample test data reported in section 7.

Questions

Re section 5 and algorithm 1, the motivation is the existence of an NN predictor of singularity from the weight matrix. Did anything else from the NN results enter into finding this result?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

yes

Reviewer Ukzu6/10 · confidence 4/52023-07-07

Summary

This paper applies machine learning methods to predict whether an algebraic variety is Q-Fano. Q-fano varieties are Fano varieties that are Q-factorial and have terminal singularities. Checking terminally is computational challenging. To train a machine learning classifier, the authors build a balanced dataset of 10 million eight-dimensional Q-factorial Fano toric varieties of Picard rank 2. They use an MLP to classify whether a topic variety has terminal singularities (and thus a Q-fano variety) from its weight matrix. The neural network achieves an accuracy of 95%, which leads to the speculation that these exists a simple criterion to check terminal singularities directly from the weight matrix. The authors develops a new algorithm that is 15 times faster than the naive approach but still orders of magnitude slower than the ML classifier. They then conclude that the ML results suggest that a simpler criterion exists. Facilitated by the faster ML model, the authors create and visualize a dataset of 100 million examples, which provides new insights for the patterns in the classification.

Strengths

- This paper provides an excellent example that shows how ML can be used to guide research in mathematics. The authors use neural networks to accelerate the classification of Fano-varieties. The computation speed by neural networks motivates new classification algorithms and suggests room for even faster ones. The trained ML model is then used as a surrogate to generate large amount of data which would be prohibitively expensive using traditional methods, and provide insights that guides theoretical understandings. While ML for science is getting popular in recent years, ML for mathematics is still a new field but has immense potential. - The Q-factorial Fano dataset provided in this paper is a valuable assets for both math and ML communities. It will advance the understanding of Q-Fano classification and benefit future development on theorem discovery. - The last section provides a number of interesting future directions. By relating the limitations of the current method to fundamental questions in ML (out-of-distribution generalization, data efficiency, geometric deep learning, interpretability), the authors demonstrate the importance of these topics and paves way for a new area in ML research.

Weaknesses

- The mathematical objects studied in this paper is not very accessible to people without a math background. As a result, the significance of results, such as the analysis of invariant of Fano-varieties section 6, is not as strong as it could be. This is perhaps due to the large amount of prerequisite for algebraic geometry, but the authors could include more introductory materials, for example, in the appendix. - While the AI-assisted workflow is novel, the contribution on machine learning is limited. In particular, the neural network uses multilayer perceptron, and no justification is given for the architecture choice. The theoretical result in section 5 may not be interesting to a NeurIPS audience. - The claim that “the ML results suggests that a simpler criterion exists” for Q-Fano classification (line 82) is not well supported. The ML model does not have 100% accuracy and does not generalize when weights matrices fall outside of the range used in training. Hence the provided results does not prove that the ML model has found a simpler and correct criterion.

Questions

- Would it be possible for the authors to comment more on the significance of the classification of Q-Fano varieties? - Does detecting terminal singularities have other applications beyond a criterion for Q-Fano varieties?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

The authors point out limitations in Section 7. There are no potential negative societal impacts of the work.

Reviewer 9Hdp7/10 · confidence 3/52023-07-15

Summary

The paper describes an ML assisted approach to the classification of Q-Fano varieties. Checking whether the singularities of a given Q-Factorial Toric variety is very challenging by mathematically established methods, involving intensive combinatorial computations. This paper applies supervised machine learning to this problem, training a feedforward NN classifier to detect terminal varieties on a balanced dataset of 5 million 8-dimensional Q-Factorial Fano toric varieties of Picard rank 2 (there exists already a fast combinatorial formula for detecting terminality in Picard rank 1). The obtained classifier obtains a 95% test accuracy on a test set of 5 million samples from the same distribution as the training data. This is surprising because the classifier is much more efficient than known combinatorial techniques, and uses only the weight matrix corresponding to the variety. The good test performance can thus be interpreted to indicate new mathematical techniques to discover. Taking inspiration from this, the second contribution of the paper is to develop an algorithm that detects terminality, directly from the weight matrix, for Q-Factorial Fano toric varieties of Picard rank two. This is the first algorithm that uses only the weight matrix in the case of Picard rank 2, and is measured to be ~15 times faster than existing methods. The new algorithm however, remains slower than the classifier. Additionally, weight information that is used by the new algorithm is unavailable to the NN classifier, having been destroyed by the encoding and regularization. This indicates that more mathematical discovery is possible. Finally, the trained classifier is used as a tool to generate a dataset of "probably terminal" examples, i.e. examples that are predicted to be terminal by the classifier. This generation is much more efficient than rigorously checking terminality. A dataset of 100M probably-Q-Fano toric varieties are then investigated by performing an $\mathbb{R}^2$ projection based on the regularized quantum period. This leads to some interesting observations about the distribution of such varieties, with the authors particularly remarking on the shape of the support and the stratification by Fano index.

Strengths

The paper makes a compelling and original case for the utility of machine learning in pure mathematics. It highlights that the performance of a classifier can be an indicator of the existence of undiscovered mathematical structure: the setting here is particularly interested due to the large discrepancy between the prediction and known mathematics. It also highlights the value of an efficient classifier for data generation that may allow for numerical investigations into the underlying structures. The new algorithm for detecting terminality is of possible additional interest. The paper also does a good job of highlighting the differences between the classifier and rigorous techniques (including the new algorithm developed here). There is also a clear discussion of the current limitations of ML assisted mathematics in this setting and general directions for future improvement.

Weaknesses

I do not see any glaring weaknesses. The approach as currently presented has limitations: the data requirements and poor OOD performance being the primary ones, but these are openly addressed and to be expected in early works in the field. The poor OOD performance is in my opinion the biggest obstacle to the claims of the paper, raising the troubling possibility that the better performance of the classifier comes from some 'finite size effect' due to the particular weight bound.

Questions

Is it reasonable to expect that a NN classifier trained and tested on a dataset with larger weight bounds would retain the surprising performance noted in the paper? Numerically probing this may be too onerous, but if the authors could describe some intuition it would help my understanding.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

3 good

Contribution

4 excellent

Limitations

Yes

Reviewer Ukzu2023-08-15

Thank you for the detailed response. I appreciate the clarifications on the significance of the Q-Fano varieties classification problem and have increased my score.

Reviewer VZKS2023-08-16

Reply to rebuttal

Thanks, this all makes sense! Some of your discussion on Q2 would be a nice addition to the paper (unless it's already there and I missed it). I confirm my rating.

Reviewer B8mu2023-08-19

ack

Thank you for the explanations.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC