DO-AutoEncoder: Learning and Intervening Bivariate Causal Mechanisms in Images

Some fundamental limitations of deep learning have been exposed such as lacking generalizability and being vunerable to adversarial attack. Instead, researchers realize that causation is much more stable than association relationship in data. In this paper, we propose a new framework called do-calculus AutoEncoder(DO-AE) for deep representation learning that fully capture bivariate causal relationship in the images which allows us to intervene in images generation process. DO-AE consists of two key ingredients: causal relationship mining in images and intervention-enabling deep causal structured representation learning. The goal here is to learn deep representations that correspond to the concepts in the physical world as well as their causal structure. To verify the proposed method, we create a dataset named PHY2D, which contains abstract graphic description in accordance with the laws of physics. Our experiments demonstrate our method is able to correctly identify the bivariate causal relationship between concepts in images and the representation learned enables a do-calculus manipulation to images, which generates artificial images that might possibly break the physical law depending on where we intervene the causal system.