Towards Principled Disentanglement for Domain Generalization (original) (raw)
Abstract:A fundamental challenge for machine learning models is generalizing to out-of-distribution (OOD) data, in part due to spurious correlations. To tackle this challenge, we first formalize the OOD generalization problem as constrained optimization, called Disentanglement-constrained Domain Generalization (DDG). We relax this non-trivial constrained optimization problem to a tractable form with finite-dimensional parameterization and empirical approximation. Then a theoretical analysis of the extent to which the above transformations deviates from the original problem is provided. Based on the transformation, we propose a primal-dual algorithm for joint representation disentanglement and domain generalization. In contrast to traditional approaches based on domain adversarial training and domain labels, DDG jointly learns semantic and variation encoders for disentanglement, enabling flexible manipulation and augmentation on training data. DDG aims to learn intrinsic representations of semantic concepts that are invariant to nuisance factors and generalizable across domains. Comprehensive experiments on popular benchmarks show that DDG can achieve competitive OOD performance and uncover interpretable salient structures within data.
Submission history
From: Hanlin Zhang [view email]
[v1] Sat, 27 Nov 2021 07:36:32 UTC (11,749 KB)
[v2] Thu, 13 Jan 2022 21:23:57 UTC (11,341 KB)
[v3] Tue, 29 Mar 2022 05:34:27 UTC (4,624 KB)
[v4] Wed, 19 Oct 2022 04:26:12 UTC (4,624 KB)