Gradient Harmony: Unifying Domain Generalization through Invariant Fusion

In order to solve the problem of image classification under different data distribution, this paper introduces a novel Principle Invariant Gradient Fusion Loss function aimed at enhancing domain generalization. By incorporating foundational principles and leveraging gradient inner products, we improve model accuracy across diverse datasets. Our approach builds upon the Empirical Risk Minimization (ERM) framework, proposing that the influence feature matrix of final decisions should align closely with the average matrix corresponding to each class. We design a principle invariant gradient loss function to reduce correlations among irrelevant features and introduce a clone model that fuses gradients during training. This fusion aligns the dominant gradient direction with those exhibiting higher inner product values. Our method addresses the challenge of learning invariant features across multiple source domains. By employing a rationale matrix to represent features, we ensure consistency across classes, enhancing the model's generalization capabilities. Our analysis demonstrates that effectively managing feature representations and gradient interactions significantly boosts model performance across varying data distributions, ultimately leading to more robust machine learning models applicable to a wide range of tasks.