Causal Intervention For Robust Classification
Hao Luo · 2024
Deep Neural Networks (DNNs) have demonstrated outstanding performance in visual recognition tasks, particularly in image classification. However, the remarkable performance of DNNs is contingent upon the training and testing datasets being drawn from the same distribution (i.e., I.I.D datasets). When the testing dataset differs in distribution from the training dataset (O.O.D datasets), the model's predictive accuracy significantly decreases. In light of this, we integrate causal reasoning and propose domain-specific causal graphs to identify the reasons behind this phenomenon. The causal graph highlights the presence of confounding connections forming a backdoor path, which leads to the model establishing spurious correlations. Employing intervention (do) reasoning, we employ the backdoor criterion to sever the backdoor paths and, in conjunction with Randomized Controlled Trials (RCTs), implement intervention reasoning. This ensures that the model learns genuine causal features for discernment without being influenced by confounding factors, thereby enhancing the model's generalization ability. We achieve outstanding performance on the out-of-distribution dataset of imagenet-9.