Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Context-Aware Self-Adaptation for Domain Generalization

Hao Yan, Yuhong Guo

TL;DR

The paper tackles domain generalization by introducing Context-Aware Self-Adaptation (CASA), a two-stage framework that simulates meta-generalization across meta-source/meta-target domain pairs. A lightweight, context-aware adaptation module CaFiLM sits between a fixed pre-trained feature extractor and classifier, using mini-batch feature means as context to adapt features while preserving performance on source domains. Training optimizes both adaptation to meta-target data and preservation of source-domain accuracy, then performs inference with an ensemble of multiple adapted meta-source models to boost robustness. Experiments on five DG benchmarks show state-of-the-art gains on smaller datasets and competitive results on larger ones, demonstrating the practical impact of context-aware adaptation and ensemble strategies for improving cross-domain generalization.

Abstract

Domain generalization aims at developing suitable learning algorithms in source training domains such that the model learned can generalize well on a different unseen testing domain. We present a novel two-stage approach called Context-Aware Self-Adaptation (CASA) for domain generalization. CASA simulates an approximate meta-generalization scenario and incorporates a self-adaptation module to adjust pre-trained meta source models to the meta-target domains while maintaining their predictive capability on the meta-source domains. The core concept of self-adaptation involves leveraging contextual information, such as the mean of mini-batch features, as domain knowledge to automatically adapt a model trained in the first stage to new contexts in the second stage. Lastly, we utilize an ensemble of multiple meta-source models to perform inference on the testing domain. Experimental results demonstrate that our proposed method achieves state-of-the-art performance on standard benchmarks.

Context-Aware Self-Adaptation for Domain Generalization

TL;DR

The paper tackles domain generalization by introducing Context-Aware Self-Adaptation (CASA), a two-stage framework that simulates meta-generalization across meta-source/meta-target domain pairs. A lightweight, context-aware adaptation module CaFiLM sits between a fixed pre-trained feature extractor and classifier, using mini-batch feature means as context to adapt features while preserving performance on source domains. Training optimizes both adaptation to meta-target data and preservation of source-domain accuracy, then performs inference with an ensemble of multiple adapted meta-source models to boost robustness. Experiments on five DG benchmarks show state-of-the-art gains on smaller datasets and competitive results on larger ones, demonstrating the practical impact of context-aware adaptation and ensemble strategies for improving cross-domain generalization.

Abstract

Domain generalization aims at developing suitable learning algorithms in source training domains such that the model learned can generalize well on a different unseen testing domain. We present a novel two-stage approach called Context-Aware Self-Adaptation (CASA) for domain generalization. CASA simulates an approximate meta-generalization scenario and incorporates a self-adaptation module to adjust pre-trained meta source models to the meta-target domains while maintaining their predictive capability on the meta-source domains. The core concept of self-adaptation involves leveraging contextual information, such as the mean of mini-batch features, as domain knowledge to automatically adapt a model trained in the first stage to new contexts in the second stage. Lastly, we utilize an ensemble of multiple meta-source models to perform inference on the testing domain. Experimental results demonstrate that our proposed method achieves state-of-the-art performance on standard benchmarks.

Paper Structure

This paper contains 16 sections, 9 equations, 1 figure, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Proposed Method
  4. Two-Stage Generalizable Learning
  5. Context-Aware Self-Adaptation Module
  6. Choice of Context Information
  7. Module Structure Design
  8. Inference with Ensemble Mechanism
  9. Experiments
  10. Experimental settings
  11. Results
  12. Ablation study
  13. Is adaptation training necessary?
  14. Is context information necessary?
  15. Self-adaptation module structure: CaFiLM or MLP?
  16. ...and 1 more sections

Figures (1)

  • Figure 1: Diagram of the proposed two-stages Context-Aware Self-Adaptation method for domain generalization. Left upper: One model is trained for each meta-source domain. Left lower: Context-Aware Self-Adaptation module is trained to adapt the pre-trained meta-source model to the meta-target domain while preserving the prediction ability on the meta-source domain. Right: Ensemble of the adapted meta-source models are used for testing.