Impact of rotational symmetry in model design

Ascertain whether incorporating rotational symmetry (e.g., C4-equivariant group convolutions or symmetric quanvolutions) systematically benefits LArTPC pixel-level track-versus-shower classification across model sizes, and characterize the regimes in which symmetry yields performance gains.

Background

The authors extend quanvolutional neural networks to include rotational symmetry beyond standard translation equivariance and compare symmetry-aware quantum and classical models.

Their results suggest that rotation symmetry appears to help only large models, and they explicitly note that its impact requires further exploration to be fully understood.