Equivariance Through Parameter-Sharing
This work addresses the challenge of designing equivariant neural networks for machine learning applications, offering a theoretical framework and practical methods, but it appears incremental as it builds on existing concepts of symmetry and parameter-sharing.
The authors tackled the problem of achieving equivariance in deep neural networks by linking it to parameter symmetries, showing that a network layer is equivariant to a group action if and only if the group explains the parameter symmetries, and they proposed two parameter-sharing schemes that ensure equivariance and sensitivity to other groups under certain conditions.
We propose to study equivariance in deep neural networks through parameter symmetries. In particular, given a group $\mathcal{G}$ that acts discretely on the input and output of a standard neural network layer $φ_{W}: \Re^{M} \to \Re^{N}$, we show that $φ_{W}$ is equivariant with respect to $\mathcal{G}$-action iff $\mathcal{G}$ explains the symmetries of the network parameters $W$. Inspired by this observation, we then propose two parameter-sharing schemes to induce the desirable symmetry on $W$. Our procedures for tying the parameters achieve $\mathcal{G}$-equivariance and, under some conditions on the action of $\mathcal{G}$, they guarantee sensitivity to all other permutation groups outside $\mathcal{G}$.