A Generalisation to other groups

The residual pathway of Eq. (10) was originally in Finzi et al. [2021a] for general groups Similar to the original residual pathway paper [Finzi et al., 2021a], we could also write linear Further, we give an extension for group convolutional layers. First, we split up the definition from Eq. (5): y ( c For sparsified factored layers S-FC, we extend the K-FAC approximation for factored layers. Similarly to above, we use the existing derivation to approximate KFAC in terms of Jacobians w.r.t. Table 4: Analysis of learned layer types after training FC+CONV model on CIFAR-10. Prior precisions learned for the CONV+FC model are shown on the left in Table 4.