Understanding Self-Training for Gradual Domain Adaptation
This addresses the challenge of adapting machine learning models to evolving data over time, such as in sensor networks or self-driving cars, with incremental theoretical and algorithmic improvements.
The paper tackles the problem of gradual domain adaptation, where a classifier must adapt to slowly shifting data distributions without labeled target data, and proves the first non-vacuous error bound for self-training, showing it outperforms direct adaptation in scenarios with small Wasserstein-infinity distance shifts.
Machine learning systems must adapt to data distributions that evolve over time, in applications ranging from sensor networks and self-driving car perception modules to brain-machine interfaces. We consider gradual domain adaptation, where the goal is to adapt an initial classifier trained on a source domain given only unlabeled data that shifts gradually in distribution towards a target domain. We prove the first non-vacuous upper bound on the error of self-training with gradual shifts, under settings where directly adapting to the target domain can result in unbounded error. The theoretical analysis leads to algorithmic insights, highlighting that regularization and label sharpening are essential even when we have infinite data, and suggesting that self-training works particularly well for shifts with small Wasserstein-infinity distance. Leveraging the gradual shift structure leads to higher accuracies on a rotating MNIST dataset and a realistic Portraits dataset.