Unsupervised Domain Adaptation with Progressive Domain Augmentation

In a world where vast amounts of common data lack the necessary labels, domain adaptation addresses this shortcoming by leveraging available labeled data from a related source domain to aid in the classification of a label-scarce target domain. In this context, unsupervised domain adaptation emerges as an essential tool, especially when the divergence between the source and target domains is notably significant.

This article introduces an innovative method that uses progressive domain augmentation for unsupervised domain adaptation. Rather than addressing the mismatch between the source and target domains directly, the method generates intermediate virtual domains via interpolation, gradually augmenting the source domain, bridging the domain divergence through multiple subspace alignment.

1. Introduction

In traditional models, training instances are assumed to be drawn from the same probability distribution as testing instances. However, in the case of domain adaptation tasks, the two domains are assumed to be running different probability distribution functions. The mismatch, often referred to as domain divergence or covariate shift, presents the challenge of learning a model that performs well on the target domain by transferring relevant knowledge learned from the source domain.

2. Related Work

Unsupervised domain adaptation methods have seen massive development. Broadly they fall under divergence-based methods, adversarial-based methods, and subspace-based methods. Data interpolation, based on augmentation techniques, offers the possibility of generating new samples and diversifying data sets. This information opens up interesting implementations, some of which the article explores.

3. Method

The proposed method, Progressive Domain Augmentation (PrDA), introduces a novel approach for unsupervised domain adaptation. Through a series of steps, a sequence of intermediate virtual domains is generated, resulting in a gradual progression of the source domain towards the target domain. The virtual samples, once generated, are incorporated into the source domain, thereby progressively moving the subspaces of the source domain closer to the target domain. Additionally, alignment of the augmented source domain and the target domain is conducted after the two domains become much closer.

The unique approach provided by PrDA handles domain adaptation with large inter-domain variations subtly, avoiding any potential information loss that traditional methods might encounter. Experiments conducted on multiple domain adaptation tasks show remarkable results, with the proposed strategy achieving state-of-the-art performance.

In conclusion, the proposed method of Unsupervised Domain Adaptation with Progressive Domain Augmentation offers a robust path towards bridging the domain divergence, presenting significant improvements in performance.