Supplementary Material Searching for Actions on the Hyperbole

Different from Nickel et al. [4], which updates the prototypes sparsely, i.e., updating one column of P per iteration, we update prototypes P globally. The reason that we don’t update sparsely is that P and its gradients are small enough to fit into one GPU’s memory. For constructing the hyperbolic space, we use geoopt1 library, which is equipped with Dc , R, Sn−1 and corresponding distance metrics, for hyperbolic manifold implementation.