🤖 machine learning

RECON: Robust symmetry discovery via Explicit Canonical Orientation Normalization

RECON introduces a class-pose agnostic canonical orientation normalization method that corrects arbitrary canonical representations via simple right translation to enable unsupervised discovery of instance-specific symmetries, detect out-of-distribution poses, and improve pre-trained models through a plug-and-play test-time layer without retraining.

Original authors: Alonso Urbano, David W. Romero, Max Zimmer, Sebastian Pokutta

Published 2026-05-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Alonso Urbano, David W. Romero, Max Zimmer, Sebastian Pokutta

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 to recognize objects, like a cat or a molecule. The problem is that these objects can appear in the real world in many different orientations: a cat might be sitting, standing, or upside down; a molecule might be rotated in 3D space.

Most AI models struggle with this. They might think a cat upside down is a completely different animal, or they might get confused when a molecule spins. To fix this, scientists usually try to force the AI to learn "symmetry" (the idea that a cat is a cat no matter how it's turned). But real-world data is messy. We don't always know exactly how things are rotated, and different objects might have different rules for how they can turn.

This paper introduces RECON, a new tool that helps AI figure out these hidden rotation rules on its own, without needing a human to label every single example.

Here is the breakdown of how it works, using simple analogies:

1. The Problem: The "Arbitrary" Photo Album

Imagine you have a photo album of handwritten numbers (like the digits 0–9).

The Old Way: An AI tries to organize these photos. It picks a "standard" version of the number "7" to represent the class. But because the AI learns randomly, it might decide that the "standard" 7 is actually tilted 45 degrees to the left.
The Result: Every time the AI sees a normal, straight 7, it thinks, "Oh, this is a 7 that has been rotated 45 degrees to the right." If it sees a 7 tilted 45 degrees left, it thinks, "This is the standard 7!"
The Confusion: The AI creates a chaotic map. It thinks the "natural" position of a 7 is tilted, and it gets confused when it sees a straight one. This makes it bad at recognizing new, unseen angles.

2. The Solution: RECON (The "Straightener")

The authors created RECON (Robust Symmetry Discovery via Explicit Canonical Orientation Normalization). Think of RECON as a smart photo editor that fixes the "tilted standard" problem.

Here is how it works in three steps:

Step A: Grouping the Similar (The "Clustering")

First, RECON looks at all the images and groups together the ones that look like the same object (e.g., all the "7"s), ignoring how they are currently rotated. It uses a special "invariant" lens that sees the shape but ignores the spin.

Step B: Finding the "Center of Gravity" (The "Freckle Mean")

Once it has a group of "7"s, it looks at all their different rotations.

Imagine you have a bunch of arrows pointing in different directions.
The old methods might just pick one random arrow as the "standard."
RECON calculates the Fréchet Mean. In simple terms, this is like finding the "average direction" of all those arrows. If the "7"s are mostly upright with a little bit of wobble, RECON finds the exact upright position that represents the center of that wobble.

Step C: The "Right Translation" (The "Correction")

This is the magic trick. RECON realizes, "Hey, the AI's original 'standard' 7 was tilted. But the real natural position is upright."
It applies a simple mathematical correction (a "right translation") to shift everything. It effectively says: "Let's rotate the entire group so that our calculated 'center' becomes the new 'upright' position."

The Result: Suddenly, all the "7"s are aligned with their natural, upright position. The AI can now clearly see that the "7"s usually appear upright with a small range of wobble (e.g., ±30 degrees), rather than being confused by an arbitrary tilt.

3. What RECON Can Do (The Superpowers)

The paper claims RECON offers three main benefits, which they tested on images (like MNIST digits) and 3D molecules:

Discovering Hidden Rules: It can look at a pile of unlabeled data and figure out, "Oh, these molecules usually spin around this specific axis," or "These digits are usually upright but can tilt a bit." It finds the "natural pose" of the object.
Spotting the Weirdos (Out-of-Distribution Detection): Because RECON knows what the "natural" range of rotation is, it can instantly spot when something is weird. If a "7" is upside down (which is outside the normal range), RECON flags it as an anomaly. It's like a security guard who knows exactly how a person should stand and immediately notices if someone is standing on their head.
The "Plug-and-Play" Upgrade: This is a major claim. You can take an AI model that was already trained (and is "frozen" or locked), and attach RECON to the front of it. RECON will rotate the incoming images to the "natural" position before the AI sees them.
- Analogy: Imagine you have a security camera trained only to see people standing straight. If you put a smart mirror (RECON) in front of it that automatically straightens anyone who is leaning, the camera suddenly works perfectly on leaning people too. You don't have to retrain the camera; you just add the mirror.

Summary

Real-world data is messy and rotated. Old AI methods often pick a "standard" view that is accidentally tilted, causing confusion. RECON is a tool that automatically finds the true, natural "center" of an object's orientation and straightens everything out. This allows AI to understand symmetry better, spot weird angles, and work much better on pre-trained models without needing to be retrained from scratch.

The authors tested this on 2D images (digits) and complex 3D molecules, showing it works well in both flat and deep spaces.

Technical Summary: RECON – Robust Symmetry Discovery via Explicit Canonical Orientation Normalization

1. Problem Statement

Real-world data often exhibits instance-specific symmetries (e.g., rotations of molecules or digits) that do not perfectly align with a pre-fixed transformation group $G$ . While class-pose decomposition methods (such as Invariant-Equivariant Autoencoders, or IE-AEs) aim to disentangle inputs into invariant class features and a pose $g \in G$ , they suffer from a critical limitation: the learned canonical representation (the reference pose) is arbitrary and training-dependent.

This arbitrariness leads to two main issues:

Out-of-Distribution (OOD) Canonicals: The canonical pose may not correspond to any natural pose observed in the data.
Arbitrary Distribution Shifts: The relative pose distributions learned by these models are shifted by the arbitrary canonical orientation, obscuring the true, instance-specific symmetry distributions. This makes it difficult to discover the natural symmetries of the data or to apply these representations effectively at test time without retraining.

Existing methods for symmetry discovery often require supervision, learn only dataset-level patterns, or fail to provide a clear, interpretable distribution for each individual instance.

2. Methodology: RECON

The authors propose RECON (Robust Symmetry Discovery via Explicit Canonical Orientation Normalization), a framework that corrects arbitrary canonicals to yield natural, data-aligned canonicalizations. The method operates in three main stages:

A. Class-Pose Decomposition Backbone

RECON leverages existing class-pose decomposition methods (specifically IE-AEs) which map an input $x$ to:

An invariant latent vector $z = \eta(x)$ .
An arbitrary canonical representation $\hat{x} = \delta(z)$ .
A relative group transformation (pose) $g = \psi(x)$ that maps $\hat{x}$ back to $x$ .

B. Equivalence Class Definition

To discover instance-specific symmetries without labels, RECON defines an equivalence class $[x]$ based on the proximity of invariant features in the latent space $Z$ . Objects with similar intrinsic shapes map to nearby points in $Z$ , forming connected components that represent the same underlying class.

C. Canonical Orientation Normalization

The core contribution is the correction of the arbitrary canonical offset.

Estimation of Offset: For a given class $[x]$ , the method collects the relative transformations $\psi([x])$ . It estimates the offset induced by the arbitrary canonical by computing the Fréchet mean (or a robust Tukey-Fréchet mean for multimodal distributions) of these transformations, denoted as $\hat{\Gamma}[x]$ .
Normalization: The method applies a right translation to the relative transformations: $\psi'(x) = \psi(x) \cdot \hat{\Gamma}[x]^{-1}$ .
Result: This operation centers the distribution of symmetries at the group identity $e$ , effectively recovering the "natural" pose distribution $\mu[x]$ where the Fréchet mean is $e$ .

D. Learnable Mappings for Inference

To enable efficient inference on unseen data without explicit class computation, RECON trains two learnable maps:

$\Theta(x)$ : Predicts the parameters of the symmetry distribution for an input.
$\Lambda(x)$ : Predicts the centering transformation (the offset) for an input.
These maps are trained using pseudo-labels generated via the normalization process described above.

3. Key Contributions

Instance-Specific Symmetry Discovery: The paper proposes a method to discover pose distributions for individual instances from unlabeled data, moving beyond dataset-level symmetry learning.
Canonical Orientation Normalization: The authors introduce a theoretically grounded correction (Proposition 3.1) that removes the bias of arbitrary canonicals. They prove that normalizing relative poses via the empirical Fréchet mean approximates the true symmetry distribution centered at the identity.
Plug-and-Play Test-Time Canonicalization: The method provides a layer that can be attached to any pre-trained model to infuse group invariance without retraining the backbone.
Applications: The framework enables unsupervised discovery of instance-specific pose distributions, detection of out-of-distribution poses, and performance improvements in downstream tasks.

4. Experimental Results

The authors validate RECON on 2D image benchmarks (MNIST, FashionMNIST) and large-scale 3D molecular data (GEOM-QM9).

Symmetry Recovery: RECON successfully recovers the ground-truth symmetry distributions (e.g., uniform rotations for specific digits, Gaussian rotations for FashionMNIST classes, and Matrix-Fisher distributions for molecules). Visualizations show that RECON produces upright, data-aligned canonicals, whereas baselines like IE-AE produce arbitrary orientations.
Partial Symmetries: Unlike fully equivariant methods that collapse distinct classes (e.g., digits '6' and '9') into the same representation, RECON distinguishes them by learning distinct offsets, correctly identifying that they are related by a $180^\circ$ rotation but possess different natural poses.
Downstream Performance:
- OOD Detection: RECON achieves higher AUC-ROC scores in detecting out-of-distribution poses compared to uncorrected IE-AE baselines.
- Test-Time Canonicalization: When applied as a pre-processing layer to frozen pre-trained models (e.g., ResNet18, GCN), RECON significantly improves classification accuracy on rotated test sets.
  - On FashionMNIST, RECON (81.96%) outperforms both the baseline (57.14%) and specialized equivariant architectures like SE(2)-Partial G-CNNs (67.72%).
  - On MNIST, it matches or approaches the performance of fully equivariant Steerable CNNs (90.96% vs 94.59% for SE(2)-ESCNN) without requiring the backbone to be equivariant.
  - On GEOM-QM9, it improves the baseline GCN performance (55.92% vs 52.40%), demonstrating effectiveness in 3D domains.

5. Significance and Claims

The paper claims that RECON addresses the "arbitrary canonical" problem inherent in class-pose decomposition, which has previously hindered the interpretability and utility of these representations. By explicitly normalizing the canonical orientation, RECON yields natural, data-aligned canonicalizations that are robust and interpretable.

The authors emphasize that their approach is architecture-agnostic and retrain-free for downstream models. It allows pre-trained models to gain group invariance simply by inserting a canonicalization layer, making it a practical "plug-and-play" solution. The method is particularly significant for 3D data (molecular conformations), where symmetry discovery is challenging and existing methods often fail to handle instance-specific variations without supervision.

The paper acknowledges limitations, noting that RECON relies on the quality of the underlying class-pose backbone and may struggle with inter-class feature overlap in complex natural images (e.g., distinguishing similar clothing items). However, it positions RECON as a foundational step toward robust, unsupervised symmetry discovery and a practical tool for improving model robustness against pose variations.