Meta-Learning Symmetries by Reparameterization

Allan Zhou, Tom Knowles, Chelsea Finn

Many successful deep learning architectures are equivariant to certain transformations in order to conserve parameters and improve generalization: most famously, convolution layers…