Regularity in Canonicalized Models: A Theoretical Perspective

Behrooz Tahmasebi, Stefanie JegelkaAISTATS 2025 · OpenReview

Abstract

In learning with invariances (or symmetries), canonicalization is a widely used technique that projects data onto a smaller subset of the input space to reduce associated redundancies. The transformed dataset is then processed through a function from a designated function class to obtain the final invariant representation. Although canonicalization is often simple and flexible, both theoretical and empirical evidence suggests that the projection map can be discontinuous and unstable, which poses challenges for machine learning applications. However, the overall end-to-end representation can still remain continuous. Focusing on the importance of end-to-end regularity rather than the projection mapping itself, this paper explores the continuity and regularity of canonicalized models from a theoretical perspective. In a broad setting of input spaces and group actions, we establish necessary and sufficient conditions for the continuity or regularity of canonicalized models of any order, thereby characterizing the minimal conditions required for stability. To our knowledge, this represents the first comprehensive investigation into the end-to-end regularity of canonicalized models, offering critical insights into their design and application, as well as guidance for enhancing stability in practical settings.

Introduction

In machine learning, achieving invariance to specific transformations—such as permutation invariance for point clouds or translation invariance for images—is a fundamental challenge. A popular approach to this problem is canonicalization. This technique involves projecting data into a "standard" or "canonical" form (e.g., sorting a list of points) before feeding it into a standard neural network. While computationally efficient compared to methods like group averaging, canonicalization faces a significant mathematical hurdle: the projection step is often inherently discontinuous. A minor perturbation in the input can lead to a drastic jump in the canonical form, potentially causing instability in the model and its gradients.

This research addresses a critical question: Can the end-to-end model (the composition of the projection and the base function) remain smooth despite the jumpy nature of the projection itself? If so, what constraints must be placed on the base function to ensure this stability?

The Canonicalization Framework

The study formalizes the canonicalization process through a topological and group-theoretic lens. The model is decomposed into three primary components:

  1. The Projection ($\pi$): Maps input data $x$ from the input space $\mathcal{X}$ to a quotient space or a canonical form.
  2. The Base Function ($f$): A function (typically a neural network) from a class $\mathcal{F}$ applied to the canonical data.
  3. The Canonicalized Model ($f \circ \pi$): The resulting end-to-end representation.
Schematic of a canonicalized model where an input $x$ is mapped by a projection $\pi$ and then processed by a base function $f$.

By identifying "critical points" where the projection map $\pi$ is prone to discontinuities, the authors derive the necessary conditions for regularity.

Theoretical Findings: Continuity and Smoothness

The paper establishes two landmark theorems that characterize the stability of these models:

  • Theorem 1 (Continuity): For the end-to-end model to be continuous, the base function $f$ must produce identical outputs for all points that the projection map "confuses" at its boundaries. Essentially, $f$ must compensate for the jumps in $\pi$.
  • Theorem 3 (Smoothness): This theorem extends the analysis to $k$-th order differentiability. It establishes that smoothness requires specific linear constraints on the derivatives of the base function at the boundaries of the canonical space.

These findings reveal a "Continuity Paradox." While canonicalization is often chosen to simplify invariant learning, ensuring the model's continuity often requires the base function itself to be invariant over significant portions of the space. This creates a circular dependency where the "easy" solution requires solving the "hard" problem it was meant to avoid.

Case Studies and Implications

The authors validated their theoretical framework using several common symmetries:

  • Sorting (Permutation Invariance): For point clouds, a canonicalized model is only differentiable if the base function is essentially already symmetric, suggesting that "smooth sorting" is highly restrictive.
  • Sign Invariance: Ensuring continuity here is mathematically equivalent to solving a sign-invariance problem on a lower-dimensional space, offering no real shortcut over traditional methods.
  • Tori (Translation Invariance): In these cases, the number of constraints needed to ensure continuity can grow exponentially with data dimensionality, sometimes requiring an infinite number of functional equalities.

Conclusion and Future Work

This work provides the first comprehensive theoretical analysis of the regularity of canonicalized models. The results highlight a fundamental trade-off: to make a canonicalized model stable enough for reliable training, the base neural network must be so heavily restricted that it may lose the expressivity and flexibility that made the approach attractive.

Future research directions include extending this analysis from invariant models to equivariant models (where outputs change predictably with inputs) and investigating the "generalization gap"—why some canonicalized models perform well in practice despite these theoretical hurdles. This study serves as a vital guide for researchers designing stable and efficient symmetry-aware machine learning systems.

Summary updated

Abstract

In learning with invariances (or symmetries), canonicalization is a widely used technique that projects data onto a smaller subset of the input space to reduce associated redundancies. The transformed dataset is then processed through a function from a designated function class to obtain the final invariant representation. Although canonicalization is often simple and flexible, both theoretical and empirical evidence suggests that the projection map can be discontinuous and unstable, which poses challenges for machine learning applications. However, the overall end-to-end representation can still remain continuous. Focusing on the importance of end-to-end regularity rather than the projection mapping itself, this paper explores the continuity and regularity of canonicalized models from a theoretical perspective. In a broad setting of input spaces and group actions, we establish necessary and sufficient conditions for the continuity or regularity of canonicalized models of any order, thereby characterizing the minimal conditions required for stability. To our knowledge, this represents the first comprehensive investigation into the end-to-end regularity of canonicalized models, offering critical insights into their design and application, as well as guidance for enhancing stability in practical settings.
BibTeX
@inproceedings{tahmasebi2025regularity,
  author = {Behrooz Tahmasebi and Stefanie Jegelka},
  title = {{Regularity in Canonicalized Models: A Theoretical Perspective}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2025},
  url = {https://openreview.net/forum?id=tBcxoC05At},
}
Figure