Equivariance with Learned Canonicalization Functions

Properties
authors	Sékou-Oumar Kaba, Arnab Kumar Mondal, Yan Zhang, Yoshua Bengio, Siamak Ravanbakhsh

Main idea

We learn a canonicalization function \(h\) either by a neural network or an optimization procedure.
$$
\phi(x) = h'(x) f(h(x)^{-1} x)
$$

Abstract