Source-linked AI summary

Domain Generalization via Model-Agnostic Learning of Semantic Features

Qi Dou, Daniel C. Castro, Konstantinos Kamnitsas, Ben Glocker

arXiv:1910.13580v1cs.CV

TL;DR

Domain generalization asks how to train on multiple source domains so models work on unseen target domains whose statistics are unknown. MASF uses model-agnostic episodic meta-training and meta-testing together with global and local semantic-feature losses, achieving state-of-the-art benchmark results and consistent medical-segmentation improvement.

  • Problem

    Domain generalization seeks models that generalize to unseen domains without access to target-domain observations or explicit target-distribution knowledge during training.

  • Method

    MASF simulates domain shift through episodic meta-train and meta-test updates, while globally aligning class relationships and locally enforcing domain-independent class-specific feature cohesion and separation.

  • Results

    MASF achieves state-of-the-art results on two object-recognition benchmarks and consistent improvement on a medical image-segmentation task.

  • Takeaways & Limitations

    Semantic feature-space constraints provide a generalization approach that improves performance across object-recognition benchmarks and medical image segmentation.

  • Takeaways & Limitations

    The domain-generalization setting assumes domain-invariant patterns exist in the inputs and can be extracted for prediction across seen and unseen domains.

Abstract

from arXiv · show

Generalization capability to unseen domains is crucial for machine learning models when deploying to real-world conditions. We investigate the challenging problem of domain generalization, i.e., training a model on multi-domain source data such that it can directly generalize to target domains with unknown statistics. We adopt a model-agnostic learning paradigm with gradient-based meta-train and meta-test procedures to expose the optimization to domain shift. Further, we introduce two complementary losses which explicitly regularize the semantic structure of the feature space. Globally, we align a derived soft confusion matrix to preserve general knowledge about inter-class relationships. Locally, we promote domain-independent class-specific cohesion and separation of sample features with a metric-learning component. The effectiveness of our method is demonstrated with new state-of-the-art results on two common object recognition benchmarks. Our method also shows consistent improvement on a medical image segmentation task.

1 Introduction

Domain generalization trains on multiple source domains to generalize directly to unseen domains without target-domain information. MASF addresses this challenge with episodic learning and complementary global and local semantic-feature regularization.

  • Unlike domain adaptation, domain generalization cannot use labelled or unlabelled target samples during training.
  • Domain generalization seeks models that generalize directly from multi-domain source data to unseen domains without target-domain information.
  • The central challenge is learning task-discriminative features that remain insensitive to domain-specific statistical changes.
  • MASF uses model-agnostic episodic learning to expose optimization to domain shift, while globally aligning class relationships and locally clustering class-specific features.
  • MASF reports state-of-the-art performance on two object-recognition benchmarks and consistent improvement on medical image segmentation.

2 Related Work

Prior domain-generalization methods learn transferable representations through domain alignment, feature regularization, and episodic meta-learning. MASF combines these directions in a model-agnostic framework summarized by episodic training, global class-relationship alignment, and local sample clustering.

  • Prior approaches align domains in input, feature, or output spaces using discrepancy metrics or adversarial learning.
  • Domain-generalization methods seek task-specific, domain-invariant representations because target-domain distributions are unavailable during training.
  • Episodic meta-learning simulates domain shift by splitting available source domains into meta-train and meta-test sets.
  • MASF’s overview combines episodic training with global alignment of class relationships and local sample clustering for cohesion and separation.

3 Method

MASF uses episodic meta-learning to expose optimization to domain shift while regularizing semantic feature structure globally and locally. It updates the model using supervised meta-train loss, then evaluates semantic objectives on held-out domains to improve generalization.

  • 3.1 Model-Agnostic Learning with Episodic Training: MASF randomly splits source domains into meta-train and meta-test sets, simulating domain shift during episodic optimization.The model is first updated on the meta-train task loss, then its semantic properties are assessed on held-out domains.
  • 3.1 Model-Agnostic Learning with Episodic Training: The meta-objective combines global class alignment and local sample clustering with positive weights β1 and β2.The resulting parameters are updated using the task loss plus the meta-objective, while the embedding network is optimized with the local loss.
  • 3.2 Global Class Alignment Objective: Global alignment matches per-class soft confusion matrices across meta-train and meta-test domains using averaged symmetrized KL divergence.Soft labels encode inter-class relationships, including class ambiguities that hard-label training does not require the model to preserve.
  • 3.3 Local Sample Clustering Objective: Local metric learning clusters same-class features and separates different-class features across domains, complementing global alignment.The objective targets domain-independent class-specific cohesion and separation to reduce cluster overlap and ambiguous decision boundaries.
  • 3.3 Local Sample Clustering Objective: Contrastive loss is used for mild domain shift, whereas triplet loss is proposed for more extreme shifts with similar semantics but distinct low-level statistics.The sample pairs are drawn across all source domains so updated features are encouraged to harmonize class-specific clustering across domains.

4 Experiments

Experiments evaluate MASF across VLCS and PACS object-recognition benchmarks, deeper PACS architectures, and multi-site brain-MRI segmentation. MASF improves over DeepAll and produces more semantically structured features across domain-shifted settings.

  • VLCS Dataset: 74.11% object-recognition accuracy on VLCS improves over DeepAll’s 72.19% average and establishes state-of-the-art performance on the benchmark.The comparison uses four-domain evaluation with one domain held out for testing.
  • PACS Dataset: 3.51% higher average accuracy than the baseline is achieved by MASF across PACS domains, including a 6.20% improvement on the unseen sketch domain.PACS uses leave-one-domain-out evaluation across art painting, cartoon, photo, and sketch.
  • PACS Dataset: MASF explicitly combines episodic domain-shift simulation with global class alignment and local sample clustering in the PACS ablation study.The ablation evaluates combinations of episodic meta-learning, global class alignment, and local sample clustering.
  • Feature-space analysis: t-SNE analysis indicates better class separation for MASF, while its embeddings show tightly grouped same-class samples separated from other classes in the silhouette analysis.The visual analyses compare MASF with DeepAll and examine class-specific cohesion and separation.
  • Deeper architectures: MASF improves over DeepAll across all PACS splits when using both ResNet-18 and ResNet-50 feature extractors.These experiments test whether the method remains beneficial with deeper architectures.
  • Tissue Segmentation in Multi-site Brain MRI: In brain-tissue segmentation across clinical centers, domain shift degrades performance by up to 10%, while MASF consistently improves over naive multi-source aggregation, especially on Set-D.The task segments background, GM, WM, and CSF in T1-weighted MRI acquired from four clinical centers.

5 Conclusions

The approach improves domain generalization by combining global and local constraints to learn semantic feature spaces. It achieves state-of-the-art results on benchmarks and improves medical image segmentation.

  • The method incorporates global and local constraints to learn semantic feature spaces for domain generalization.
  • It achieves new state-of-the-art results on popular benchmarks and a dense medical image classification task, semantic segmentation.
  • The proposed loss functions are generally orthogonal to other algorithms, motivating future evaluation of their integration.
Loading 1910.13580v1…