Efficiently Approximating Equivariance in Unconstrained Models

Equivariance ML models has emerged as a powerful inductive bias for leveraging symmetries in 3D data, achieving significant success across different scientific applications in molecular modeling and protein generation. Despite these advances, strictly enforcing equivariance can impose prohibitive computational costs. Meanwhile, recent breakthroughs in protein structure prediction—exemplified by AlphaFold 3’s reliance on data augmentation techniques—have demonstrated that unconstrained architectures can still yield outstanding results, culminating in a Nobel Prize for its transformative impact on 3D protein modeling.

In this talk, we will introduce a new training procedure that approximates equivariance via a simple multitask objective. By adding an additional equivariance loss to unconstrained models, our approach enables these architectures to learn approximate symmetries without incurring the overhead of fully equivariant methods. Crucially, formulating equivariance as a flexible learning objective allows precise control over the extent to which symmetry is enforced, matching the performance of strictly equivariant baselines at a fraction of the training and inference cost.

We will conclude by surveying key achievements, current challenges, and future directions for equivariant models, emphasizing the balance between rigorous theory and practical scalability in geometric deep learning.