CMGAN: A generative adversarial network embedded with causal matrix

Wenbin Zhang, Jun Liao, Yi Zhang, Li Liu

Abstract: Exploring the relationship between different features in multi-label learning and generating controllable features of multi-label images have always been hot topics. Unlike most other methods, in order to make the processing of the network more intuitive and convincing, we designed a generative adversarial network with a nested structural causal mechanism. The network has two functions: generating feature labels and generating multi-label images. We restore the adjacency matrix between individual features in the labels by applying an additional network layer with acyclic constraints, so as to obtain the corresponding causal directed acyclic graph. In this way, the network can autonomously learn the causal relationships between features and generate feature labels from them, enabling feature separation at the causal level. A particular innovation is that we can also add intervention mechanism to the label generation process to answer the “what if” question in the causality theory at the feature level. Finally, the second part of the network maps the labels to images, completing the entire process of multi-label image generation. We performed a series of experiments to demonstrate the effectiveness of the embedded causal mechanism in qualitative and quantitative terms.