Manifold GPLVMs for discovering non-Euclidean latent structure in neural data