Source-linked AI summary
Cross-Domain Few-Shot Classification via Adversarial Task Augmentation
Haoqing Wang, Zhi-Hong Deng
TL;DR
Cross-domain few-shot classification remains difficult because inductive biases learned from source tasks may not generalize under domain shift. The paper introduces adversarial task augmentation to generate inductive-bias-adaptive challenging tasks, and reports improved cross-domain performance across meta-learning models. Its scope focuses on single-domain generalization with fixed task sample counts.
Problem
Meta-learning inductive biases can fail to generalize when training and test tasks come from different domains, degrading few-shot classification performance.
Method
The method solves a worst-case problem around source task distribution T0 to generate challenging, inductive-bias-adaptive virtual tasks as a plug-and-play module.
Results
The method improves cross-domain few-shot classification for RelationNet, GNN, and TPN, outperforming existing works and traditional pre-training and fine-tuning under domain shift.
Takeaways & Limitations
Adversarial task augmentation provides a general task-level approach for improving the cross-domain robustness of various meta-learning models.
Takeaways & Limitations
The setting focuses on single-domain generalization with unknown target domains and assumes each task has a fixed number of samples.
Abstract
from arXiv · showhide
Few-shot classification aims to recognize unseen classes with few labeled samples from each class. Many meta-learning models for few-shot classification elaborately design various task-shared inductive bias (meta-knowledge) to solve such tasks, and achieve impressive performance. However, when there exists the domain shift between the training tasks and the test tasks, the obtained inductive bias fails to generalize across domains, which degrades the performance of the meta-learning models. In this work, we aim to improve the robustness of the inductive bias through task augmentation. Concretely, we consider the worst-case problem around the source task distribution, and propose the adversarial task augmentation method which can generate the inductive bias-adaptive 'challenging' tasks. Our method can be used as a simple plug-and-play module for various meta-learning models, and improve their cross-domain generalization capability. We conduct extensive experiments under the cross-domain setting, using nine few-shot classification datasets: mini-ImageNet, CUB, Cars, Places, Plantae, CropDiseases, EuroSAT, ISIC and ChestX. Experimental results show that our method can effectively improve the few-shot classification performance of the meta-learning models under domain shift, and outperforms the existing works. Our code is available at https://github.com/Haoqing-Wang/CDFSL-ATA.
1 Introduction
The paper addresses cross-domain few-shot classification, where meta-learning inductive biases learned on source tasks may fail under domain shift. It proposes adversarial task augmentation that generates challenging, inductive-bias-adaptive tasks and improves cross-domain generalization.
- Meta-learning models perform undesirably when training and test tasks come from different domains, sometimes underperforming traditional pre-training and fine-tuning.
- The paper seeks a general, easy-to-implement method that improves the robustness of various inductive biases through adaptive task augmentation.
- The method considers a worst-case problem around the source task distribution T0 and generates challenging tasks within a wider distribution space.
- The method is presented as the first task-augmentation approach for improving cross-domain few-shot classification under domain shift.
- It is a plug-and-play inductive bias-adaptive module for various meta-learning models.
- Extensive cross-domain experiments with RelationNet, GNN, and TPN show improved generalization, outperforming existing works and traditional pre-training and fine-tuning.
2 Related Work
Related work has sought to improve cross-domain generalization through feature transformations or explanation-guided feature reweighting, but these approaches target particular components or require model-specific explanations.
- Feature-wise transformation layers simulate varied image-feature distributions during training to improve a metric function’s generalization.
- Explanation-based methods emphasize prediction-relevant features and penalize them more, providing another route to cross-domain generalization.
- Explanation-guided training requires manually deriving explanations for different meta-learning models.
3 Method
The method formulates adversarial task augmentation as a worst-case optimization around the source task distribution, generating challenging tasks adapted to each model’s inductive bias. It uses gradient-based task updates, random convolutions, and a plug-and-play training procedure to improve robustness under domain shift.
- Preliminaries: Few-shot tasks contain support and query sets, and meta-learning models learn task-shared inductive bias from tasks sampled from a source distribution.The base learner uses the support set to produce a task-specific classifier, which is evaluated on the query set.
- Cross-Domain Setting: The cross-domain setting trains on one source domain and requires generalization to several unknown target domains.The method focuses on single-domain generalization, where multiple training domains may be unavailable because of acquisition cost or privacy constraints.
- Adversarial Task Augmentation: The method solves a worst-case problem over task distributions near the source distribution using Wasserstein distance to allow different task supports.The optimization uses a Lagrangian relaxation and conditions for a unique worst-case task under smoothness and strong-convexity assumptions.
- Adversarial Task Augmentation: The generated virtual task is more challenging than the source task and adapts to the inductive bias through the model-dependent loss.The resulting task is used to update model parameters, with the method implemented as a plug-and-play module.
- Adversarial Task Augmentation: For complex models, gradient ascent with early stopping updates task samples while keeping labels fixed and preserving the task’s sample count.After Tmax iterations, the resulting task is paired with the original labels and used for parameter updates.
- Adversarial Task Augmentation: Random convolution changes local textures while preserving image shape and is applied uniformly to task samples with probability 1−p.This auxiliary augmentation is intended to address CNN sensitivity to superficial local textures.
4 Experiments
Experiments evaluate adversarial task augmentation across diverse cross-domain few-shot settings and meta-learning models. The method consistently improves generalization, while regularization terms provide no clear benefit.
- Evaluation: Experiments use RelationNet, GNN, and TPN to test the method across different inductive biases.The evaluation compares these models with feature-wise transformation and explanation-guided training.
- Experimental Settings: The benchmark covers nine datasets, using mini-ImageNet as the source domain and eight other domains for evaluation.The reported settings include 5-way 1-shot/5-shot tasks with average accuracy measured across sampled episodes.
- Main Results: Adversarial task augmentation consistently and significantly improves cross-domain few-shot accuracy across the evaluated meta-learning models.It achieves comparable or greater improvements than feature-wise transformation and explanation-guided training, and remains effective for TPN.
- Ablation Study: Random convolution is not necessary for improvement, but including it produces further gains.The ablation compares the complete method with a version without random convolution across eight unseen domains.
- Ablation Study: Neither sample-wise Euclidean distance nor MMD regularization provides obvious benefits and can instead harm performance.With γ = 1 and random convolution disabled, the results suggest early stopping already supplies sufficient constraint.
- Comparison with Fine-Tuning: The method can outperform traditional pre-training and fine-tuning under domain shift when evaluated under the same settings.The comparison uses augmented support data for fine-tuning on unseen tasks.
5 Conclusion
The paper introduces a plug-and-play adversarial task augmentation method for improving meta-learning under cross-domain few-shot learning. It reports stronger cross-domain performance than existing methods and fine-tuning under domain shift.
- Method: The method generates challenging virtual tasks through an inductive bias-adaptive task augmentation module.It is designed as a plug-and-play approach for various meta-learning models.
- Results: The approach significantly improves cross-domain few-shot classification for multiple meta-learning models.The conclusion describes the method as improving the cross-domain generalization capability of meta-learning models.
- Results: Under domain shift, meta-learning models equipped with the method outperform fine-tuning.The comparison is conducted under the same settings.