Exploring dropout discriminator for domain adaptation

Adaptation of a classifier to new domains is one the challenging problems in machine learning. This has been addressed using many deep and non-deep learning based methods. Among methodologies used, that adversarial widely applied solve along with domain adaptation. These methods are on discriminator ensures source target distributions close. However, here we suggest rather than point estimate obtaining by single discriminator, it would be useful if distribution ensembles discriminators could used bridge this gap. achieved multiple classifiers or traditional ensemble In contrast, Monte Carlo dropout suffice obtain discriminator. Specifically, propose curriculum gradually increases variance sample corresponding reverse gradients align feature representations. An helps model learn data efficiently. It also provides better gradient estimates train extractor. The detailed results thorough ablation analysis show our outperforms state-of-the-art results.