Improving Clustering Uncertainty-weighted Embeddings for Active Domain Adaptation

Domain adaptation generalizes deep neural networks to new target domains under domain shift. Active domain adaptation (ADA) does so efficiently by allowing the learning model to strategically ask data annotation questions. The state-of-the-art active domain adaptation via clustering uncertaintyweighted embeddings (ADA-CLUE) uses uncertainty-weighted clustering to identify target instances for labeling. In this work, we carefully study how ADA-CLUE balances uncertainty and diversity during active learning. We compare the original ADA-CLUE with a variant that weights clusters by a constant instead of by the uncertainty, and confirm that constant-weighted clustering sampling outperforms ADA-CLUE at early stages due to its stability. We then merge constant-weighted sampling and uncertainty-weighted sampling with a threshold to get the best of the two worlds. The merged solution, called CLUE with a loop threshold, is shown to be an empirically better choice than the original ADA-CLUE.