Paper ID: 2304.01973

ERM++: An Improved Baseline for Domain Generalization

Piotr Teterwak, Kuniaki Saito, Theodoros Tsiligkaridis, Kate Saenko, Bryan A. Plummer

Domain Generalization (DG) measures a classifier's ability to generalize to new distributions of data it was not trained on. Recent work has shown that a hyperparameter-tuned Empirical Risk Minimization (ERM) training procedure, that is simply minimizing the empirical risk on the source domains, can outperform most existing DG methods. ERM has achieved such strong results while only tuning hyper-parameters such as learning rate, weight decay, batch size, and dropout. However there are additional hyperparameters which further limit overfitting and catastrophic forgetting. We therefore focus on tuning previously untuned hyper-parameters, including training amount, initialization, and additional regularizers. We call the resulting stronger baseline ERM++. ERM++ improves the performance of DG by over 5% compared to prior ERM baselines on a standard benchmark of 5 datasets with a ResNet-50 and over 15% with a ViT-B/16, and outperforms all SOTA methods on DomainBed with both architectures. We also explore the relationship between DG performance and similarity to pre-training data, and find that similarity to pre-training data distributions is an important driver of performance, but that ERM++ with stronger initializations can deliver strong performance even on dissimilar datasets.Code is released at https://github.com/piotr-teterwak/erm_plusplus.

Submitted: Apr 4, 2023