In deep learning, symmetry is often a double-edged sword. We want our models to be "equivariant"—meaning if an input rotates, scales, or shifts, the output should change in a predictable, corresponding way. Traditionally, achieving this required building specialized architectures from scratch, using mathematically complex layers like group convolutions or steerable filters.
However, a powerful alternative has emerged: Modular Canonicalization. Instead of baking symmetry into the internal layers of a network, researchers are wrapping standard, high-performance backbones in a "canonicalization" layer. This approach decouples the task of handling geometry from the task of extracting features, allowing us to retrofit massive, pre-trained models with strict geometric guarantees.
The Canonicalize-then-Process Paradigm
The core intuition behind this research direction is simple: if you want to recognize an object regardless of its orientation, it is easier to rotate the object to a "standard" (canonical) pose before looking at it than it is to learn what that object looks like from every possible angle.
Sékou-Oumar Kaba, Siamak Ravanbakhsh, and their colleagues formalized this in Equivariance with Learned Canonicalization Functions. They proposed using a lightweight, equivariant "predictor" network to estimate the transformation needed to move data into a reference frame. Once the data is aligned, a standard, non-equivariant backbone—like a ResNet or a Transformer—processes it. Because the backbone always sees the data in the same orientation, the entire system becomes equivariant by construction.
This "wrapper" strategy is highly modular. As Ngoc Thieu Vo demonstrates in Design equivariant neural networks for 3D point cloud, this allows researchers to use optimized, general-purpose architectures (like MLPs or standard convolutions) as the "engine" of an equivariant system, avoiding the need to design specialized layers from scratch.
Retrofitting Foundation Models
One of the most significant advantages of modular equivariance is its ability to "upgrade" existing large-scale models. Foundation models like the Segment Anything Model (SAM) or Vision Transformers (ViT) are incredibly powerful but are not inherently equivariant to rotations or flips.
Arnab Kumar Mondal and Sékou-Oumar Kaba showed in Equivariant Adaptation of Large Pretrained Models that adding a tiny canonicalization network (adding as little as 0.3% more parameters) can make these models robust to transformations without losing their original performance. Similarly, Sourya Basu and Lav R. Varshney introduced $\lambda$-equitune in Equivariant Few-Shot Learning from Pretrained Models, a post-hoc wrapping method that allows models like GPT-2 to handle symmetries in reinforcement learning and language tasks without expensive retraining.
From Global to Local Frames
While global canonicalization works for rigid objects, it struggles with complex, articulated shapes like a human body or a molecule. To solve this, the research has moved toward Local Canonicalization.
In Equivariant Point Cloud Analysis via Learning Orientations for Message Passing, Shitong Luo and colleagues suggest learning a unique coordinate frame for every single point in a cloud. By projecting neighbors into these local frames, the model decouples local geometry from global pose.
Peter Lippmann and Gerrit Gerhartz took this further in Beyond Canonicalization: How Tensorial Messages Improve Equivariant Message Passing. They showed that by transforming "tensorial messages" between these local frames, standard message-passing networks can achieve exact $O(d)$-equivariance. This same principle was applied to high-energy physics by Jonas Spinner and Luigi Favaro in Lorentz Local Canonicalization, allowing any standard backbone to respect the laws of special relativity by processing particle data in locally canonicalized Lorentz frames.
Why It Works: Decoupling and Efficiency
The success of modular canonicalization stems from three key factors:
- Architectural Freedom: By handling symmetry externally, researchers can use the most modern, non-equivariant components—like Softmax attention or LayerNorm—which are often mathematically incompatible with internal equivariance constraints. Youssef Saied and François Fleuret demonstrate this in Normalization Equivariance for Arbitrary Backbones, proving that a simple "normalize-process-denormalize" wrapper is the mathematically complete way to achieve scale and shift robustness for any backbone.
- Computational Speed: Specialized equivariant layers are often computationally heavy. In contrast, canonicalization wrappers are lightweight. Luigi Favaro and Fred A Hamprecht found that their LLoCa framework was up to 10x faster than specialized equivariant models while maintaining the same physical symmetries.
- Data Efficiency: Instead of using data augmentation to "teach" a model that a rotated chair is still a chair, these frameworks enforce that knowledge through the geometry of the wrapper. As shown by Matan Atzmon and Yaron Lipman in Frame Averaging for Equivariant Shape Space Learning, this leads to perfect generalization to unseen orientations, even with limited training data.