Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Mixstyle-Entropy: Domain Generalization with Causal Intervention and Perturbation

Luyao Tang, Yuxuan Yuan, Chaoqi Chen, Xinghao Ding, Yue Huang

TL;DR

This work tackles domain generalization by reframing learning through a causal lens. It introduces InPer, a framework that combines entropy-based causal intervention during training (EnIn) with homeostatic-score guided causal perturbation during testing (HoPer), grounded in a structural causal model to disentangle domain-related signals from semantically relevant features. EnIn selectively perturbs high-entropy, domain-informative features to isolate causal content, while HoPer constructs a prototype classifier from causal-stable test samples and uses perturbation-based refinement constrained by HomeoScore. Across multi-source classification and semantic segmentation benchmarks, InPer achieves state-of-the-art results, with ablations confirming the contributions of EnIn and HoPer and practical guidance on patch size and testing protocols. The approach offers a plug-and-play path to robust DG with minimal additional parameters, highlighting the practical impact of incorporating causal interventions into both training and testing pipelines.

Abstract

Despite the considerable advancements achieved by deep neural networks, their performance tends to degenerate when the test environment diverges from the training ones. Domain generalization (DG) solves this issue by learning representations independent of domain-related information, thus facilitating extrapolation to unseen environments. Existing approaches typically focus on formulating tailored training objectives to extract shared features from the source data. However, the disjointed training and testing procedures may compromise robustness, particularly in the face of unforeseen variations during deployment. In this paper, we propose a novel and holistic framework based on causality, named InPer, designed to enhance model generalization by incorporating causal intervention during training and causal perturbation during testing. Specifically, during the training phase, we employ entropy-based causal intervention (EnIn) to refine the selection of causal variables. To identify samples with anti-interference causal variables from the target domain, we propose a novel metric, homeostatic score, through causal perturbation (HoPer) to construct a prototype classifier in test time. Experimental results across multiple cross-domain tasks confirm the efficacy of InPer.

Mixstyle-Entropy: Domain Generalization with Causal Intervention and Perturbation

TL;DR

This work tackles domain generalization by reframing learning through a causal lens. It introduces InPer, a framework that combines entropy-based causal intervention during training (EnIn) with homeostatic-score guided causal perturbation during testing (HoPer), grounded in a structural causal model to disentangle domain-related signals from semantically relevant features. EnIn selectively perturbs high-entropy, domain-informative features to isolate causal content, while HoPer constructs a prototype classifier from causal-stable test samples and uses perturbation-based refinement constrained by HomeoScore. Across multi-source classification and semantic segmentation benchmarks, InPer achieves state-of-the-art results, with ablations confirming the contributions of EnIn and HoPer and practical guidance on patch size and testing protocols. The approach offers a plug-and-play path to robust DG with minimal additional parameters, highlighting the practical impact of incorporating causal interventions into both training and testing pipelines.

Abstract

Despite the considerable advancements achieved by deep neural networks, their performance tends to degenerate when the test environment diverges from the training ones. Domain generalization (DG) solves this issue by learning representations independent of domain-related information, thus facilitating extrapolation to unseen environments. Existing approaches typically focus on formulating tailored training objectives to extract shared features from the source data. However, the disjointed training and testing procedures may compromise robustness, particularly in the face of unforeseen variations during deployment. In this paper, we propose a novel and holistic framework based on causality, named InPer, designed to enhance model generalization by incorporating causal intervention during training and causal perturbation during testing. Specifically, during the training phase, we employ entropy-based causal intervention (EnIn) to refine the selection of causal variables. To identify samples with anti-interference causal variables from the target domain, we propose a novel metric, homeostatic score, through causal perturbation (HoPer) to construct a prototype classifier in test time. Experimental results across multiple cross-domain tasks confirm the efficacy of InPer.
Paper Structure (13 sections, 11 equations, 4 figures, 3 tables)

This paper contains 13 sections, 11 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Causality and Domain Generalization
  4. Online Parameter Updating
  5. Methodology
  6. Theoretical Insight
  7. Training: Entropy Based Casual Intervention
  8. Testing: Homeostatic Score Based Causal Perturbation
  9. Experiments
  10. Generalization on Multi-Source Domain Classification
  11. Generalization on Semantic Segmentation
  12. Analysis and Ablations
  13. Conclusions

Figures (4)

  • Figure 1: The pipeline of the proposed InPer. Training: Extract visible entropy and perform causal interventions on samples to sever the spurious connections between class and domain. Testing: Causal perturbation is applied to the testing samples, and through the HomeoScore, a generalized classifier is built to adapt to the target domain.
  • Figure 2: The visualization on unseen domain Cityscapes with the model trained on GTA5.
  • Figure 3: The t-SNE visualization of extracted deep features using different methods on PACS dataset (target domain: cartoon). The different colors stands for different classes.
  • Figure 4: (a) and (b) are pseudo-label selection capability between entropy and HomenScore. (c) is effects of patch sizes. (d) is the influence of test batch size in different schemes.