[Paper Review] Open Domain Generalization with Domain-Augmented Meta-Learning

Shu Yang, Zhangjie Cao+3 more|arXiv (Cornell University)|Apr 8, 2021

Domain Adaptation and Few-Shot Learning55 references8 citations

TL;DR

This paper proposes Domain-Augmented Meta-Learning (DAML), a novel framework for Open Domain Generalization (OpenDG) that addresses distribution shifts and disparate label sets across source domains. By introducing Dirichlet mixup for feature-level augmentation and distilled soft-labeling for label-level knowledge transfer, DAML enables meta-learning across augmented domains, achieving state-of-the-art performance on unseen target domains with significantly improved generalization on both known and open classes.

ABSTRACT

Leveraging datasets available to learn a model with high generalization ability to unseen domains is important for computer vision, especially when the unseen domain's annotated data are unavailable. We study a novel and practical problem of Open Domain Generalization (OpenDG), which learns from different source domains to achieve high performance on an unknown target domain, where the distributions and label sets of each individual source domain and the target domain can be different. The problem can be generally applied to diverse source domains and widely applicable to real-world applications. We propose a Domain-Augmented Meta-Learning framework to learn open-domain generalizable representations. We augment domains on both feature-level by a new Dirichlet mixup and label-level by distilled soft-labeling, which complements each domain with missing classes and other domain knowledge. We conduct meta-learning over domains by designing new meta-learning tasks and losses to preserve domain unique knowledge and generalize knowledge across domains simultaneously. Experiment results on various multi-domain datasets demonstrate that the proposed Domain-Augmented Meta-Learning (DAML) outperforms prior methods for unseen domain recognition.

Motivation & Objective

• To address the challenge of generalizing deep models to unseen target domains with unknown distributions and label sets.
• To overcome the limitations of existing domain generalization methods that assume shared label sets or require target domain data.
• To develop a unified framework that learns generalizable representations across source domains with disparate label sets and distributions.
• To enable effective knowledge transfer across domains with minimal assumptions on label set overlap or target data availability.

Proposed method

• Proposes a novel Open Domain Generalization (OpenDG) problem setting that requires no assumptions on label sets, target data, or post-training on target data.
• Introduces Dirichlet mixup (Dir-mixup), a feature-level augmentation that generates diverse mixed samples across multiple domains using Dirichlet distribution sampling.
• Applies distilled soft-labeling to transfer knowledge from other domains’ trained networks to enrich missing class representations in each source domain.
• Designs domain-augmented meta-learning tasks and losses that preserve domain-specific knowledge while promoting cross-domain generalization.
• Uses ensemble prediction from all source domain networks to naturally calibrate uncertainty and improve robustness on open-set recognition.
• Employs meta-learning over augmented domains to learn representations that generalize to unseen domains with minimal inductive bias.

Figure 1: Open Domain Generalization (OpenDG). Different source domains hold disparate label sets. The goal is to learn generalizable representations from these source domains to help classify the known classes and detect open classes in the unseen target domain.

Experimental results

Research questions

• RQ1 Can a meta-learning framework effectively generalize across source domains with disparate label sets and distributions?
• RQ2 How can feature-level and label-level domain augmentation improve generalization in open domain generalization?
• RQ3 Does Dirichlet mixup outperform standard mixup in promoting representation diversity and generalization?
• RQ4 To what extent does distilled soft-labeling enhance model performance on minority and missing classes?
• RQ5 Can meta-learning over domain-augmented data improve zero-shot generalization to unseen target domains?

Key findings

• DAModel achieves a 56.5% average accuracy on the OpenDG benchmark, significantly outperforming prior methods such as CuMix and AGG.
• The ablation study shows that combining Dir-mixup and distilled soft-labeling with meta-learning yields the highest performance, with a 56.5% average accuracy on Office-Home.
• Using Dir-mixup instead of standard mixup improves accuracy by 1.3% on average, demonstrating the benefit of multi-domain mixup for diversity.
• The inclusion of distilled soft-labeling increases accuracy by 1.2% on average, confirming its effectiveness in transferring cross-domain knowledge.
• Meta-learning with domain-augmented data reduces the Fréchet distance between source and target domain features, indicating better domain alignment.
• The model achieves 53.1% accuracy on the open-set part of the target domain, showing strong generalization to unknown classes.

Figure 2: The architecture of the proposed DAML framework. We show the computation graph for source domain 1 as an example, and the other source domains are computed similarly. In the meta-training (up part, left to right), each source domain is augmented by Dir-mixup (red) and distilled soft-labeli