Source-linked AI summary
Heterogeneous Domain Generalization via Domain Mixup
Yufei Wang, Haoliang Li, Alex C. Kot
TL;DR
The paper addresses heterogeneous domain generalization, where a DCNN feature extractor must support novel categories in an unseen target domain with different label spaces. It proposes mixing samples across source domains using two sampling strategies and reports that the resulting method outperforms existing baselines on the Visual Decathlon benchmark.
Problem
Heterogeneous domain generalization seeks a feature extractor that supports recognition of novel categories in a novel target domain despite disjoint source and target label spaces.
Method
The method constructs an HDG mixup distribution by combining samples between two domains and across multiple domains.
Results
The proposed methods outperform existing state-of-the-art baselines in classification accuracy and VD-score on the Visual Decathlon benchmark.
Takeaways & Limitations
Domain mixup produces a more robust feature extractor that can be used off-the-shelf for heterogeneous domain generalization.
Takeaways & Limitations
The HDG setting assumes disjoint label spaces between domains.
Abstract
from arXiv · showhide
One of the main drawbacks of deep Convolutional Neural Networks (DCNN) is that they lack generalization capability. In this work, we focus on the problem of heterogeneous domain generalization which aims to improve the generalization capability across different tasks, which is, how to learn a DCNN model with multiple domain data such that the trained feature extractor can be generalized to supporting recognition of novel categories in a novel target domain. To solve this problem, we propose a novel heterogeneous domain generalization method by mixing up samples across multiple source domains with two different sampling strategies. Our experimental results based on the Visual Decathlon benchmark demonstrates the effectiveness of our proposed method. The code is released in \url{https://github.com/wyf0912/MIXALL}
1 University of Electronic Science and Technology of China, China 2 Nanyang Technological University, Singapore
The paper concerns heterogeneous domain generalization, mixup, and generalization capability.
- The paper addresses heterogeneous domain generalization.
- The paper uses mixup as a relevant approach.
- The paper targets generalization capability.
1. INTRODUCTION
The introduction motivates heterogeneous domain generalization as a response to limited DCNN generalization and distinguishes it from related domain-shift settings. The paper proposes domain mixup across source domains and reports improved performance on the Visual Decathlon benchmark.
- DCNNs often lack generalization capability, especially when large-scale training data are unavailable.The introduction discusses pretraining, finetuning, and feature extraction as common responses.
- Domain adaptation uses target-domain samples during training, whereas domain generalization trains only on source-domain data for unseen targets.
- Heterogeneous domain generalization allows source and target domains to have different label spaces.
- The proposed method regularizes models using convex combinations of samples from different domains through two mixup algorithms.The strategies model linear behavior between training examples and linear relationships among multiple domains.
- Experiments on the Visual Decathlon benchmark report performance exceeding state-of-the-art baselines for heterogeneous domain generalization.
2. PROPOSED METHOD
The method trains a shared feature extractor for heterogeneous domains by constructing mixed feature-target pairs from source domains with disjoint label spaces. It uses domain mixup through two-domain and all-domain sampling strategies, with β controlling mixture concentration and randomness.
- Problem formulation: The HDG setting assumes source domains have different distributions and disjoint label spaces, requiring labels to be represented with one-hot encodings.The method augments labels so mixed feature-target pairs can be sampled across heterogeneous domains.
- Problem formulation: HDG trains a shared feature extractor on source domains and adapts it to target domains using shallow classifiers such as SVM or KNN.Source domains provide sufficient labeled data, while target domains have fewer labeled samples and are used for downstream evaluation.
- Mixup distribution: The original mixup distribution forms convex combinations of feature-target pairs, with λ controlling the mixing degree.The mixing coefficient is sampled from a distribution and produces a generated sample and corresponding label.
- HDG mixup distribution: HDG mixup extends mixup to data from different domains, integrating heterogeneous-domain information into sampled feature-target pairs.The distribution is designed because HDG training data come from different distributions and use disjoint label spaces.
- MIXUP from two domains: The first strategy, MIXUP, samples two data pairs from arbitrary source domains and mixes them using λ drawn from Beta(α, α).The sampled pairs are combined into a new mixed feature-target pair, with α controlling λ.
- MIXUP from all domains: MIX-ALL mixes samples from all source domains in one iteration, while β controls the expectation of the maximum mixture weight and the variance of λ.Increasing β makes generated samples more definite and less random according to the supplied description and Fig. 1.
3. EXPERIMENT
The experiments evaluate MIX-ALL on the Visual Domain Decathlon benchmark using SVM and KNN classifiers, with comparisons against established heterogeneous-domain generalization baselines. Results report stronger classification accuracy and VD-Score, while ablations examine the effects of the two mixup hyperparameters.
- Dataset and setup: The Visual Domain Decathlon benchmark contains ten heterogeneous domains, split into six source domains and four target domains for evaluation.Source domains provide training data, while target-domain features are evaluated using shallow classifiers.
- Dataset and setup: Experiments use ResNet-18 with its first four blocks frozen and evaluate extracted target-domain features using SVM and KNN.The study follows the Feature-Critic Networks setup and reports classification accuracy and VD-Score.
- Main results: MIX-ALL significantly outperforms all baselines in both classification accuracy and VD-Score when SVM is used for target-domain classification.VD-Score reflects whether performance remains stable across the benchmark.
- Main results: MIX-ALL also achieves much better performance than all baseline methods when KNN is used as the classifier.This provides a second evaluation of the learned feature representation under a different shallow classifier.
- Ablation study: For MIXUP, performance first improves and then drops as α increases, while values near zero reduce the method to the baseline.The authors associate the rise with enhanced robustness and the decline with underfitting.
- Ablation study: For MIX-ALL, average accuracy drops when β equals 0 or 1, large β values approach the baseline, and performance continues to decline as the number of domains n increases.The authors relate small β to excessive noise and large β to degeneration into the baseline.
4. CONCLUSION
The paper introduces domain mixup for heterogeneous domain generalization, using two sampling strategies to augment data across domains and produce a more robust feature extractor.
- The method applies domain mixup as data augmentation for heterogeneous domain generalization.
- It uses two strategies: mixup between two domains and mixup across multiple domains.
- The resulting robust feature extractor can be used off-the-shelf.
- The methods outperform existing state-of-the-art baselines in classification accuracy and VD-score.