From Hope to Safety: Unlearning Biases of Deep Models via Gradient Penalization in Latent Space
This addresses the risk of biased predictions in high-stake decision-making applications like medicine, offering a novel approach for model correction on the concept level.
The paper tackles the problem of deep neural networks learning spurious correlations that lead to biased predictions, presenting a method that reduces model sensitivity to biases via gradient penalization, achieving effective bias mitigation on datasets like ISIC, Bone Age, ImageNet, and CelebA using architectures such as VGG, ResNet, and EfficientNet.
Deep Neural Networks are prone to learning spurious correlations embedded in the training data, leading to potentially biased predictions. This poses risks when deploying these models for high-stake decision-making, such as in medical applications. Current methods for post-hoc model correction either require input-level annotations which are only possible for spatially localized biases, or augment the latent feature space, thereby hoping to enforce the right reasons. We present a novel method for model correction on the concept level that explicitly reduces model sensitivity towards biases via gradient penalization. When modeling biases via Concept Activation Vectors, we highlight the importance of choosing robust directions, as traditional regression-based approaches such as Support Vector Machines tend to result in diverging directions. We effectively mitigate biases in controlled and real-world settings on the ISIC, Bone Age, ImageNet and CelebA datasets using VGG, ResNet and EfficientNet architectures. Code is available on https://github.com/frederikpahde/rrclarc.