A Prototype-Oriented Framework for Unsupervised Domain Adaptation: Appendix

Since it is difficult to directly optimize the marginal likelihood due to the sum inside the log function, we resort to the Expectation-Maximization (EM) algorithm, where we iterate between the expectation and maximization steps. In practice, we draw a mini-batch of size M to estimate this quantity. In other words, we assign each data point to its closest centroid. We report the average accuracy from three independent runs. We do not perform any additional hyper-parameter searches.