[Paper Review] Improve Unsupervised Domain Adaptation with Mixup Training

Shen Yan, Huan Song+3 more|arXiv (Cornell University)|Jan 3, 2020

Domain Adaptation and Few-Shot LearningComputer Science17 references122 citations

TL;DR

The paper introduces IIMT, a framework that enforces cross-domain and intra-domain mixup training with a feature-level consistency regularizer and domain adversarial learning to improve unsupervised domain adaptation across vision and HAR tasks.

ABSTRACT

Unsupervised domain adaptation studies the problem of utilizing a relevant source domain with abundant labels to build predictive modeling for an unannotated target domain. Recent work observe that the popular adversarial approach of learning domain-invariant features is insufficient to achieve desirable target domain performance and thus introduce additional training constraints, e.g. cluster assumption. However, these approaches impose the constraints on source and target domains individually, ignoring the important interplay between them. In this work, we propose to enforce training constraints across domains using mixup formulation to directly address the generalization performance for target data. In order to tackle potentially huge domain discrepancy, we further propose a feature-level consistency regularizer to facilitate the inter-domain constraint. When adding intra-domain mixup and domain adversarial learning, our general framework significantly improves state-of-the-art performance on several important tasks from both image classification and human activity recognition.

Motivation & Objective

• Address the limitation of domain-adversarial methods that treat source and target constraints separately.
• Leverage mixup to create cross-domain supervision and enforce linearity in predictions.
• Introduce a feature-level consistency regularizer to aid mixup under large domain discrepancy.
• Demonstrate improvements over state-of-the-art methods on visual and HAR domain adaptation benchmarks.

Proposed method

• Apply inter-domain mixup between source and target with inferred target labels as weak supervision.
• Compute target virtual labels by augmenting target samples, averaging predictions, and sharpening with a temperature-controlled softmax.
• Enforce cross-domain label-level mixup loss and a cross-domain feature-level mixup loss (L_q and L_z).
• Incorporate a domain adversarial loss to encourage domain-invariant features (DANN).
• Perform intra-domain mixup within source and within target domains using label-level mixup (L_s and L_t).
• Combine all losses into a final objective with tunable weights and a gradually increasing weight for target-specific loss.

Experimental results

Research questions

• RQ1 Can cross-domain mixup supervision improve generalization to the unlabeled target domain beyond traditional domain-adversarial training?
• RQ2 Does a feature-level consistency regularizer facilitate effective mixup when domain discrepancy is large?
• RQ3 How do inter-domain and intra-domain mixup components contribute to performance across image classification and HAR tasks?

Key findings

• IIMT significantly outperforms state-of-the-art methods on several visual UDA benchmarks and HAR tasks.
• In ablations, inter-domain and intra-domain mixup components each provide gains, with combined inter-domain mixup yielding substantial improvements.
• Intra-domain mixup alone can rival other Lipschitz-constraint methods like VAT in stabilizing predictions.
• The framework demonstrates robust improvements across diverse domains, including MNIST, SVHN, CIFAR- STL, and OPPORTUNITY HAR datasets.