Purifying Confounders in Vision Transformer via Causal Intervention

Causal intervention has surfaced as a method to aid deep learning models in precisely capturing authentic causal relationships among variables, gaining growing traction in various computer vision studies in recent years. Employing causal intervention aids models in removing the impact of confounders between variables, thereby preventing the acquisition of spurious correlations. However, due to the complexity associated with implementing causal intervention, current research predominantly applies it to classification tasks. In this paper, we integrate causal intervention into our Confounder-purify Vision Transformer (CP ViT). Not only do we outperform other state-of-the-art (SOTA) causal intervention methods in accuracy during ImageN et-9 classification experiments, but we also achieve the lowest error in regression experiments for Bone Age prediction. These experimental results demonstrate the successful integration of causal intervention with ViT, enabling our model to achieve confounders purification in both classification and regression tasks. This ensures a more accurate modeling of causal relationships between variables.