On the Implicit Bias of Linear Equivariant Steerable Networks

Indeed, recent studies have shown that gradient-based training methods effectively regularize the solution by implicitly minimizing a certain complexity measure of the model [V ardi, 2022].