A Appendix

Finally, when using different values for P, we can get other group actions. Let us first show that (2) and (3) correspond to a particular case of Cohen et al. This proves (2) and (3). In both subcases, by Lemma 4, θ must be a leaky ReLu function. Given a non-equivariant model, a simple way to let it "learn" to be equivariant is to train it with This doubles the size of the training set, which increases the training time.