Virtual Adversarial Training in Feature Space to Improve Unsupervised Video Domain Adaptation

Virtual Adversarial Training has recently seen a lot of success in\nsemi-supervised learning, as well as unsupervised Domain Adaptation. However,\nso far it has been used on input samples in the pixel space, whereas we propose\nto apply it directly to feature vectors. We also discuss the unstable behaviour\nof entropy minimization and Decision-Boundary Iterative Refinement Training\nWith a Teacher in Domain Adaptation, and suggest substitutes that achieve\nsimilar behaviour. By adding the aforementioned techniques to the state of the\nart model TA$^3$N, we either maintain competitive results or outperform prior\nart in multiple unsupervised video Domain Adaptation tasks\n