Robust Domain-Free Domain Generalization with Class-aware Alignment

TL;DR

This paper introduces Domain-Free Domain Generalization (DFDG), a novel, model-agnostic approach that enhances generalization to unseen domains by learning domain-invariant features without relying on source domain labels.

Contribution

The paper proposes DFDG, a new method that aligns class relationships and uses saliency maps to improve domain generalization without source domain labels.

Findings

Achieves competitive results on time series sensor datasets.

Performs well on image classification benchmarks.

Effectively learns domain-invariant features.

Abstract

While deep neural networks demonstrate state-of-the-art performance on a variety of learning tasks, their performance relies on the assumption that train and test distributions are the same, which may not hold in real-world applications. Domain generalization addresses this issue by employing multiple source domains to build robust models that can generalize to unseen target domains subject to shifts in data distribution. In this paper, we propose Domain-Free Domain Generalization (DFDG), a model-agnostic method to achieve better generalization performance on the unseen test domain without the need for source domain labels. DFDG uses novel strategies to learn domain-invariant class-discriminative features. It aligns class relationships of samples through class-conditional soft labels, and uses saliency maps, traditionally developed for post-hoc analysis of image classification networks, to remove superficial observations from training inputs. DFDG obtains competitive performance on both time series sensor and image classification public datasets.