On Image Classification: Correlation v.s. Causality

Image classification is one of the fundamental problems in computer vision. Owing to the availability of large image datasets like ImageNet and YFCC100M, a plethora of research has been conducted to do high precision image classification and many remarkable achievements have been made. The success of most existing methods hinges on a basic hypothesis that the testing image set has the same distribution as the training image set (i.e. the i.i.d. hypothesis). However, in many real applications, we cannot guarantee the validity of the i.i.d. hypothesis since the testing image set is unseen. It is thus desirable to learn an image classifier, which can perform well even in non-i.i.d. situations. In this paper, we propose a novel Causally Regularized Logistic Regression (CRLR) algorithm to address the non-i.i.d. problem without knowing testing data information by searching for causal features. The causal features refer to characteristics truly determining whether a special object belongs to a category or not. Identifying causal features allows us to construct classifiers adaptive to distributional changes in the non i.i.d circumstances even when the testing set is unseen. Algorithmically, we propose a causal regularizer for causal feature identification by jointly optimizing it with a logistic loss term. Assisted with the causal regularizer, we can estimate the causal contribution (causal effect) of each focal image feature (viewed as a treatment variable) by sample reweighting which ensures the distributions of all remaining image features between images with different focal feature levels are close. The resultant classifier will be based on the estimated causal contributions of the features, rather than traditional correlation-based contributions. To validate the effectiveness of our CRLR algorithm, we manually construct a new image dataset from YFCC100M1, simulating various non-i.i.d. situations in the real world, and conduct extensive experiments for image classification. Experimental results clearly demonstrate that our CRLR algorithm outperforms the state-of-the-art methods. We further visualize the top causal features selected by our algorithm on our image dataset.