Preprint
Representation Learning for Sequence Data with Deep Autoencoding Predictive Components
arXiv (Cornell University)
02/28/2021
DOI: 10.48550/arxiv.2010.03135
Abstract
We propose Deep Autoencoding Predictive Components (DAPC) -- a self-supervised representation learning method for sequence data, based on the intuition that useful representations of sequence data should exhibit a simple structure in the latent space. We encourage this latent structure by maximizing an estimate of predictive information of latent feature sequences, which is the mutual information between past and future windows at each time step. In contrast to the mutual information lower bound commonly used by contrastive learning, the estimate of predictive information we adopt is exact under a Gaussian assumption. Additionally, it can be computed without negative sampling. To reduce the degeneracy of the latent space extracted by powerful encoders and keep useful information from the inputs, we regularize predictive information learning with a challenging masked reconstruction loss. We demonstrate that our method recovers the latent space of noisy dynamical systems, extracts predictive features for forecasting tasks, and improves automatic speech recognition when used to pretrain the encoder on large amounts of unlabeled data.
Details
- Title: Subtitle
- Representation Learning for Sequence Data with Deep Autoencoding Predictive Components
- Creators
- Junwen Bai - Cornell UniversityWeiran Wang - University of California, MercedYingbo Zhou - SalesforceCaiming Xiong - Salesforce
- Resource Type
- Preprint
- Publication Details
- arXiv (Cornell University)
- DOI
- 10.48550/arxiv.2010.03135
- eISSN
- 2331-8422
- Language
- English
- Date posted
- 02/28/2021
- Academic Unit
- Computer Science
- Record Identifier
- 9984696573202771
Metrics
1 Record Views