Abstract
Recently Active Domain Adaptation (ADA) has seen its emergence to boost performance with actively labeled target samples. However, previous work can only deal with a single target domain yet the performance is unknown when multiple targets are available. In this project, we propose an active training and sampling framework called ??? to tackle this challenging Multi-Target Active Domain Adaptation (MT-ADA) task. In particular, we leverage the geometric properties of pseudo-gradients in high-dimensional embedding space to select samples beneficial to both classification and domain alignment. We
show ??? ’s superior performance on a variety of experiments with different labeling budgets, conducted on image domain adaptation datasets Office-31 and Office-Home.
Motivation
Nowadays, more and more deep-learning-based models are deployed on real-world tasks. However, because of the biased training data domain distribution, there would be performance degradation when testing on different domains, a.k.a domain gap. In this project, we want to bridge the gap and boost cross-domain performance in the training process.
Problem setting
Suppose that we have a single-domain labeled source dataset S with k classes and we can access N multi-domain unlabeled target datasets with the same classes. We aim at training one image classification model that performs well on all target domains. And we report the average top-1 accuracy across different source domains and all target domains as the evaluation metric.