Learning Gradient-based Mixup with Extrapolation toward Flatter Minima for Domain Generalization
Researchers have developed a new domain generalization technique called Flatness-aware Gradient-based Mixup (FGMix). This method uses data interpolation and extrapolation to improve model generalization by covering a wider range of feature space. FGMix assigns instance weights based on gradient compatibilities, aiming to learn mixup policies that lead to flatter minima and better performance on unseen domains. Experiments on the DomainBed benchmark showed FGMix outperforming existing domain generalization algorithms. AI
IMPACT Introduces a novel technique to improve model robustness against distribution shifts, potentially enhancing performance in real-world applications with varied data.