Gradca: Generalizing to Unseen Domains Via Gradient Calibration

Abstract

In domain generalization (DG) problem, we hope to train a robust model from multiple source domains and generalize it to unseen domains. However, the trained model can not perform well in the unseen domain since the domain-invariant representation is hard to learn across multiple source domains. In this work, we propose a novel gradient optimization strategy named gradient calibration (GradCa) which optimizes the dominant gradients and conflicting gradients without learning extra parameters. The dominant gradients are suppressed for addressing to be inclined to the optimization directions of special domains. Then two strategies (average and sign-mask) are designed for further alleviating the conflicting gradients. These types of gradients are effectively calibrated for improving the domain-invariant representation learning. Extension experiments on four benchmark datasets have shown the competitive results and the effectiveness of improving the model generalization with GradCa.

Keywords: Image ClassificationDomain generalizationGradient optimization