Improving sequential latent variable models with autoregressive flows
- Creators
- Marino, Joseph
- Chen, Lei
- He, Jiawei
- Mandt, Stephan
Abstract
We propose an approach for improving sequence modeling based on autoregressive normalizing flows. Each autoregressive transform, acting across time, serves as a moving frame of reference, removing temporal correlations and simplifying the modeling of higher-level dynamics. This technique provides a simple, general-purpose method for improving sequence modeling, with connections to existing and classical techniques. We demonstrate the proposed approach both with standalone flow-based models and as a component within sequential latent variable models. Results are presented on three benchmark video datasets and three other time series datasets, where autoregressive flow-based dynamics improve log-likelihood performance over baseline models. Finally, we illustrate the decorrelation and improved generalization properties of using flow-based dynamics.
Additional Information
© The Author(s), under exclusive licence to Springer Science+Business Media LLC, part of Springer Nature 2021. Received 15 November 2020; Revised 02 August 2021; Accepted 30 September 2021; Published 18 November 2021. Author Contributions: Joseph Marino: forming idea, running simulation and writing paper; Lei Chen: running simulation; Jiawei He: writing paper, and Stephan Mandt: forming idea, writing paper and providing feedback. All authors worked toegther for most of the ta sks involved. Conflict of interest: California Institute of Technology (caltech.edu); Simon Fraser University (sfu.ca); University of California Irvine (uci.edu); Disney Research (disneyresearch.com); Borealis AI (borealisai.com); DeepMind (deepmind.com, google.com). Code availability: Source code is available at https://anonymous.4open.science/r/f02199f7-86d2-45ee-ad23-3f13f769ee10/.Attached Files
Submitted - 2010.03172.pdf
Files
Name | Size | Download all |
---|---|---|
md5:2491ea960045f472584045d9740bb710
|
7.4 MB | Preview Download |
Additional details
- Eprint ID
- 112663
- DOI
- 10.1007/s10994-021-06092-6
- Resolver ID
- CaltechAUTHORS:20220104-767137600
- Created
-
2022-01-04Created from EPrint's datestamp field
- Updated
-
2022-06-01Created from EPrint's last_modified field