The Hessian Penalty: A Weak Prior for Unsupervised Disentanglement
Name
2008.10599.pdf
Description
Accepted version
Size
12.81 MB
Format
Adobe PDF
Checksum (MD5)
ac65ae40607a8a608937eba1f7ef597a
Author(s)
Peebles, William
Peebles, John
Zhu, Jun-Yan
Efros, Alexei
Torralba, Antonio
Date Issued
November 2020
Journal
Lecture Notes in Computer Science
Publisher
Springer International Publishing
Citation
Peebles, William et al. "The Hessian Penalty: A Weak Prior for Unsupervised Disentanglement."
ECCV: European Conference on Computer Vision, Lecture Notes in Computer Science, 12351, Springer International Publishing, 2020, 581-597. © 2020, Springer Nature
Version
Author's final manuscript
Abstract
Existing disentanglement methods for deep generative models rely on hand-picked priors and complex encoder-based architectures. In this paper, we propose the Hessian Penalty, a simple regularization term that encourages the Hessian of a generative model with respect to its input to be diagonal. We introduce a model-agnostic, unbiased stochastic approximation of this term based on Hutchinson’s estimator to compute it efficiently during training. Our method can be applied to a wide range of deep generators with just a few lines of code. We show that training with the Hessian Penalty often causes axis-aligned disentanglement to emerge in latent space when applied to ProGAN on several datasets. Additionally, we use our regularization term to identify interpretable directions in BigGAN’s latent space in an unsupervised fashion. Finally, we provide empirical evidence that the Hessian Penalty encourages substantial shrinkage when applied to over-parameterized latent spaces. We encourage readers to view videos of our disentanglement results at www.wpeebles.com/hessian-penalty, and code at https://github.com/wpeebles/hessian_penalty.
Description
Part of the Lecture Notes in Computer Science book series (LNCS, volume 12351)
MIT Department
Massachusetts Institute of Technology. Computer Science and Artificial Intelligence Laboratory
Terms of Use
Creative Commons Attribution-Noncommercial-Share Alike
Persistent DSpace Link
DOI of Published Version
https://doi.org/10.1007/978-3-030-58539-6_35
