Understanding Self-Training for Gradual Domain Adaptation

By Ananya Kumar, Tengyu Ma, Percy Liang. 2020

272 citations recorded by the source corpus.

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.

Subfields: Artificial Intelligence, Representation & Self-Supervised Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Representation & Self-Supervised Learning

Citations in the corpus

Break-It-Fix-It: Unsupervised Learning for Program Repair (2021, neutral): “Self-training (Lee, 2013; McClosky et al., 2006; Kumar et al., 2020; Xie et al., 2021) applies a trained model to unlabeled data, obtains predicted targets (pseudolabels), and uses them as extra training examples.”
Improving Question Answering Model Robustness with Synthetic Adversarial Data Generation (2021, methodology): “Self-training has been shown to improve robustness to domain shift (Kumar et al., 2020) and, in our case, we re-label answers to the generated questions based on the six QA model predictions.”
WILDS: A Benchmark of in-the-Wild Distribution Shifts (2020, neutral): “Utilizing this gradual shift structure with the timestamp metadata may enable adaptation across longer time periods (Kumar et al., 2020).”
Question Answering Infused Pre-training of General-Purpose Contextualized Representations (2021, methodology): “Since this student model has the same architecture as the teacher model, we train it to match the teacher’s argmax predictions, a standard self-training objective (Lee, 2013; Kumar et al., 2020).”
Data Feedback Loops: Model-driven Amplification of Dataset Biases (2022, background): “Data feedback has connections to various domain adaptation settings (Farahani et al., 2021; Shu et al., 2018; Kumar et al., 2020; Lipton et al., 2018), where the changing data distributions can be viewed as shifting target domains.”
ST++: Make Self-trainingWork Better for Semi-supervised Semantic Segmentation (2021, background): “Recently, it is increasingly attracting attention from multiple fields, such as fully-supervised image recognition [43, 53, 54, 63], semi-supervised learning [7, 17, 48, 55], and domain adaptation [30, 64, 65].”
In-N-Out: Pre-Training and Self-Training using Auxiliary Information for Out-of-Distribution Robustness (2020, methodology): “Following [27], we employ additional regularization when doing self training by adding dropout with probability 0.”
Data Selection for Language Models via Importance Resampling (2023, background): “However, in many modern applications the source and target are often disjoint (e.g., sketches vs. natural images), causing undefined importance weights [62; 42; 73].”