Current browse context:
cs.LG
Change to browse by:
References & Citations
Computer Science > Machine Learning
Title: Radon Sobolev Variational Auto-Encoders
(Submitted on 29 Nov 2019 (v1), last revised 14 Apr 2021 (this version, v3))
Abstract: The quality of generative models (such as Generative adversarial networks and Variational Auto-Encoders) depends heavily on the choice of a good probability distance. However some popular metrics like the Wasserstein or the Sliced Wasserstein distances, the Jensen-Shannon divergence, the Kullback-Leibler divergence, lack convenient properties such as (geodesic) convexity, fast evaluation and so on. To address these shortcomings, we introduce a class of distances that have built-in convexity. We investigate the relationship with some known paradigms (sliced distances - a synonym for Radon distances -, reproducing kernel Hilbert spaces, energy distances). The distances are shown to possess fast implementations and are included in an adapted Variational Auto-Encoder termed Radon Sobolev Variational Auto-Encoder (RS-VAE) which produces high quality results on standard generative datasets.
Keywords: Variational Auto-Encoder; Generative model; Sobolev spaces; Radon Sobolev Variational Auto-Encoder;
Submission history
From: Gabriel Turinici [view email][v1] Fri, 29 Nov 2019 15:02:28 GMT (1496kb,D)
[v2] Mon, 16 Mar 2020 17:00:16 GMT (1526kb,D)
[v3] Wed, 14 Apr 2021 18:08:35 GMT (1951kb,D)
Link back to: arXiv, form interface, contact.