Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AdvST: Revisiting Data Augmentations for Single Domain Generalization

Guangtao Zheng, Mengdi Huai, Aidong Zhang

TL;DR

AdvST reframes standard data augmentations as learnable semantics transformations and optimizes a mini-max objective to train models robust to unseen domain shifts from a single source domain. By interpreting the learned augmentations as semantics-induced distributions $Q_{\psi}$ and leveraging a Wasserstein-based DRO formulation, AdvST provides a tractable inner optimization and expands target-domain coverage. Empirically, AdvST and its entropy-regularized variant AdvST-ME achieve state-of-the-art average SDG performance on the Digits, PACS, and DomainNet benchmarks and remain effective with limited source data. This work demonstrates that principled, learnable augmentation semantics can effectively simulate diverse domain shifts using only the source domain, with practical implications for SDG in vision applications.

Abstract

Single domain generalization (SDG) aims to train a robust model against unknown target domain shifts using data from a single source domain. Data augmentation has been proven an effective approach to SDG. However, the utility of standard augmentations, such as translate, or invert, has not been fully exploited in SDG; practically, these augmentations are used as a part of a data preprocessing procedure. Although it is intuitive to use many such augmentations to boost the robustness of a model to out-of-distribution domain shifts, we lack a principled approach to harvest the benefit brought from multiple these augmentations. Here, we conceptualize standard data augmentations with learnable parameters as semantics transformations that can manipulate certain semantics of a sample, such as the geometry or color of an image. Then, we propose Adversarial learning with Semantics Transformations (AdvST) that augments the source domain data with semantics transformations and learns a robust model with the augmented data. We theoretically show that AdvST essentially optimizes a distributionally robust optimization objective defined on a set of semantics distributions induced by the parameters of semantics transformations. We demonstrate that AdvST can produce samples that expand the coverage on target domain data. Compared with the state-of-the-art methods, AdvST, despite being a simple method, is surprisingly competitive and achieves the best average SDG performance on the Digits, PACS, and DomainNet datasets. Our code is available at https://github.com/gtzheng/AdvST.

AdvST: Revisiting Data Augmentations for Single Domain Generalization

TL;DR

AdvST reframes standard data augmentations as learnable semantics transformations and optimizes a mini-max objective to train models robust to unseen domain shifts from a single source domain. By interpreting the learned augmentations as semantics-induced distributions and leveraging a Wasserstein-based DRO formulation, AdvST provides a tractable inner optimization and expands target-domain coverage. Empirically, AdvST and its entropy-regularized variant AdvST-ME achieve state-of-the-art average SDG performance on the Digits, PACS, and DomainNet benchmarks and remain effective with limited source data. This work demonstrates that principled, learnable augmentation semantics can effectively simulate diverse domain shifts using only the source domain, with practical implications for SDG in vision applications.

Abstract

Single domain generalization (SDG) aims to train a robust model against unknown target domain shifts using data from a single source domain. Data augmentation has been proven an effective approach to SDG. However, the utility of standard augmentations, such as translate, or invert, has not been fully exploited in SDG; practically, these augmentations are used as a part of a data preprocessing procedure. Although it is intuitive to use many such augmentations to boost the robustness of a model to out-of-distribution domain shifts, we lack a principled approach to harvest the benefit brought from multiple these augmentations. Here, we conceptualize standard data augmentations with learnable parameters as semantics transformations that can manipulate certain semantics of a sample, such as the geometry or color of an image. Then, we propose Adversarial learning with Semantics Transformations (AdvST) that augments the source domain data with semantics transformations and learns a robust model with the augmented data. We theoretically show that AdvST essentially optimizes a distributionally robust optimization objective defined on a set of semantics distributions induced by the parameters of semantics transformations. We demonstrate that AdvST can produce samples that expand the coverage on target domain data. Compared with the state-of-the-art methods, AdvST, despite being a simple method, is surprisingly competitive and achieves the best average SDG performance on the Digits, PACS, and DomainNet datasets. Our code is available at https://github.com/gtzheng/AdvST.
Paper Structure (36 sections, 2 theorems, 15 equations, 9 figures, 7 tables, 1 algorithm)

This paper contains 36 sections, 2 theorems, 15 equations, 9 figures, 7 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. AdvST: Adversarial Learning With Semantics Transformations
  4. Semantics Transformation
  5. Learning Objective of AdvST
  6. Learning Algorithm
  7. Theoretical Analysis: Connection to DRO
  8. DRO Formulation
  9. Semantics-Induced Distribution
  10. Uncertainty Set of AdvST
  11. DRO Learning Objective for AdvST
  12. Experiment
  13. Experimental Settings
  14. Ablation Studies
  15. Comparison on Digits
  16. ...and 21 more sections

Key Result

Proposition 1

Let $\ell: \Theta \times \mathcal{X} \times \mathcal{Y} \rightarrow [0,\infty)$ denote the loss function which is upper semi-continuous and integrable. The transportation cost function $c:\Xi\times \Xi \rightarrow [0,\infty)$ with $\Xi=\mathcal{X} \times \mathcal{Y}$ is a lower semi-continuous funct where $\mathcal{Q}_{\Psi}$ is a set of distributions induced by $M$ semantics transformations param

Figures (9)

  • Figure 1: Visualization of how samples from the source domain, target domains, and synthetic domains distribute in the embedding space. We compare AdvST and AdvST-ME with their non-semantics counterparts ADA and ME-ADA.
  • Figure 2: Accuracy heatmap for models trained individually on the five domains from the Digit dataset using ERM.
  • Figure 3: Average classification accuracy under different ratios of available training data.
  • Figure A4: Sensitivity analysis on $\lambda$. We train models with AdvST under different values of $\lambda$. For each $\lambda$, we report average classification accuracy (blue bars) and its standard deviation (vertical black bars) over all target domains for each dataset.
  • Figure A5: Examples of the Digits dataset.
  • ...and 4 more figures

Theorems & Definitions (3)

  • Definition 1
  • Proposition 1
  • Lemma 1