Representation Disentanglement via Regularization by Causal Identification
Primary Area: causal reasoning
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.
Keywords: Causality, disentanglement, deep learning.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: In this work, we argue modern deep representation learning models for disentanglement are ill-posed with collider bias behavior; a source of bias producing dependencies between the underlying generating variables. Under the rubric of causal inference, we show this issue can be explained and reconciled under the condition of causal identification; attainable from a combination of a causal graphical model encoding the data generation process assumptions and data. For this, we propose regularization by identification (ReI), a modular regularization engine designed to align the behavior of large scale models with the disentanglement constraints imposed by causal identification. Empirical evidence on standard disentanglement benchmarks demonstrates the superiority of ReI in removing the effects of collider-bias. In a real-world dataset we show that enforcing ReI in a variational framework results in interpretable representations robust to out-of-distribution examples and that align with the true expected effect from domain knowledge.
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: 7568
Loading