Pseudo Labels for Unsupervised Domain Adaptation: A Review

Conventional machine learning relies on two presumptions: (1) the training and testing datasets follow same independent distribution, (2) an adequate quantity of samples is essential for achieving optimal model performance during training. Nevertheless, meeting these assumptions can be challenging in real-world scenarios. Domain adaptation (DA) a subfield transfer that focuses reducing distribution difference between source domain (Ds) target (Dt) subsequently applying knowledge gained from Ds task to Dt task. The majority current DA methods aim achieve invariance by aligning marginal probability distributions Ds. Dt. Recent studies have pointed out alone not sufficient alignment conditional equally important migration. Nonetheless, unsupervised presents more significant difficulty because unavailability labels In response this issue, there been several proposed researchers, including pseudo-labeling, which offer novel solutions tackle problem. paper, we systematically analyze various pseudo-labeling algorithms their applications DA. First , summarize pseudo-label generation based single multiple classifiers actions taken deal with problem imbalanced samples. Second, investigate application category feature improving discrimination. Finally, point challenges trends algorithms. As far as know, article initial review techniques