Symbolic Autoencoding with Straight-Through Gradient Approximations
Keywords: straight-through gradient approximation, Autoencoders, discrete representation learning
TL;DR: A latent variable modeling paradigm for discrete and sequential random variables; utilized for connecting sequence-to-sequence models and training them end-to-end with unparallel data.
Abstract: Self-supervised autoregressive models have achieved significant success across diverse domains, including text,
audio, and biological sequences. However, these models often rely heavily on large samples of aligned (parallel)
data, limiting their applicability in low-resource settings.
To address this limitation, we propose Symbolic Autoencoding ($\Sigma$AE) with Straight-Through Gradient Estimators
(STGEs)—a latent variable model where the latent space consists of sequences of categorical random variables,
resembling sentences in an emergent symbolic language. $\Sigma$AE is trained end-to-end using a family of
straight-through gradient estimators.
In the unsupervised mode, $\Sigma$AE learns to compress input data into symbolic sentences and reconstructs the data
from this emergent language. In weakly supervised settings, $\Sigma$AE further grounds the latent language by
leveraging supervised training on the small amount of parallel data available.
We evaluate $\Sigma$AE with three well-known quantization mechanisms on four text sequence transduction tasks.
Our results show that $\Sigma$AE outperforms baseline methods, particularly in low-resource scenarios with limited
parallel data.
Supplementary Material: zip
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 11108
Loading