Current browse context:
stat.ML
Change to browse by:
References & Citations
Statistics > Machine Learning
Title: Spread Divergences
(Submitted on 21 Nov 2018 (v1), revised 3 Feb 2021 (this version, v4), latest version 5 Dec 2022 (v5))
Abstract: For distributions p and q with different supports, the divergence D(p|q) may not exist. We define a spread divergence on modified p and q and describe sufficient conditions for the existence of such a divergence. We demonstrate how to maximize the discriminatory power of a given divergence by parameterizing and learning the spread. We also give examples of using a spread divergence to train and improve implicit generative models, including linear models (Independent Components Analysis) and non-linear models (Deep Generative Networks).
Submission history
From: Mingtian Zhang [view email][v1] Wed, 21 Nov 2018 22:35:08 GMT (2907kb,D)
[v2] Sun, 2 Dec 2018 10:29:46 GMT (2907kb,D)
[v3] Tue, 10 Dec 2019 13:32:11 GMT (5082kb,D)
[v4] Wed, 3 Feb 2021 03:07:02 GMT (3310kb,D)
[v5] Mon, 5 Dec 2022 04:24:31 GMT (3310kb,D)
Link back to: arXiv, form interface, contact.