Hyperspherical Variational Autoencoders Using Efficient Spherical Cauchy Distribution
Researchers have introduced a new method for variational autoencoders designed for hyperspherical latent spaces, utilizing an efficient spherical Cauchy distribution. This approach offers a robust and scalable alternative for generative modeling, particularly for image and molecular sequence data. The proposed spCauchy distribution exhibits heavy-tailed behavior and allows for precise differentiable reparameterization, outperforming existing von Mises-Fisher based methods in terms of stability and evaluation speed on both CPU and GPU. AI
IMPACT Introduces a novel, more stable, and faster generative modeling technique for hyperspherical data.