Abstract
Invariance under symmetry is an important problem in machine learning. Ourpaper looks specifically at equivariant neural networks where transformationsof inputs yield homomorphic transformations of outputs. Here, steerable CNNshave emerged as the standard solution. An inherent problem of steerablerepresentations is that general nonlinear layers break equivariance, thusrestricting architectural choices. Our paper applies harmonic distortionanalysis to illuminate the effect of nonlinearities on Fourier representationsof SO(2). We develop a novel FFT-based algorithm for computing representationsof non-linearly transformed activations while maintaining band-limitation. Ityields exact equivariance for polynomial (approximations of) nonlinearities, aswell as approximate solutions with tunable accuracy for general functions. Weapply the approach to build a fully E(3)-equivariant network for sampled 3Dsurface data. In experiments with 2D and 3D data, we obtain results thatcompare favorably to the state-of-the-art in terms of accuracy while permittingcontinuous symmetry and exact equivariance.