Theory Bridges Symmetry and Classification: Optimal Group Actions for Neural Networks

A new theoretical framework connects the design of symmetry-aware neural networks to the underlying statistical structure of classification problems. The work provides principled guidance for selecting which symmetries a network should respect—potentially improving accuracy in domains where data distributions have known geometric properties.

The Problem: Choosing Symmetries Wisely#

Group-equivariant neural networks impose mathematical symmetries—constraints that ensure the model respects certain structural properties of the input space. For instance, a convolutional network preserves translation symmetry; other architectures might preserve rotations or permutations. However, practitioners have lacked a principled way to decide which symmetries to build in. Choosing the right ones can boost performance, while forcing inappropriate symmetries wastes model capacity or introduces bias.

A Theory-Driven Solution#

The researchers leverage the Neyman-Pearson lemma, a foundational result in statistical hypothesis testing, to identify optimal group actions for binary classification. By analyzing how group symmetries interact with fiber decompositions (a way to understand data structure under symmetry transformations) and likelihood ratios (the statistical signature of class differences), they construct a unified methodology. Essentially, the theory determines which symmetries align with how the true underlying probability distributions actually differ between classes.

From Theory to Practice: Jet Classification#

The framework is applied to jet classification at the Large Hadron Collider—a real-world test case where particle physics provides ground truth about symmetries. The work shows that smaller permutation symmetries among jet constituents can outperform larger or more restrictive symmetry choices, suggesting that over-constraining a model can indeed be counterproductive. This demonstrates the practical value of matching architectural symmetries to the actual data statistics rather than assuming larger symmetry groups are always better.

Why It Matters#

This contribution addresses a key gap in deep learning: moving from empirical symmetry design toward a principled, mathematics-backed approach. The methodology could inform network architecture choices across applications where domain knowledge about symmetry exists—from particle physics to chemistry to materials science—helping researchers build more efficient and accurate models.