Causal Adversarial Network for Learning Conditional and Interventional Distributions
Raha Moraffah Department of Computer Science Arizona State University Tempe, AZ 85281 \AndBahman Moraffah School of Electrical, Computer, and Energy Engineering Arizona state university Tempe AZ, 85287 \AndMansooreh Karami Department of Computer Science Arizona State University Tempe, AZ 85281 \AndAdrienne Raglin Army Research Lab 2800 Powder Mill Rd Adelphi, MD 20783 \AndHuan Liu Department of Computer Science Arizona State University Tempe, AZ 85281
Abstract
We propose a generative Causal Adversarial Network (CAN) for learning and sampling from conditional and interventional distributions. In contrast to the existing CausalGAN which requires the causal graph to be given, our proposed framework learns the causal relations from the data and generates samples accordingly. The proposed CAN comprises a two-fold process namely Label Generation Network (LGN) and Conditional Image Generation Network (CIGN). The LGN is a GAN-based architecture which learns and samples from the causal model over labels. The sampled labels are then fed to CIGN, a conditional GAN architecture, which learns the relationships amongst labels and pixels and pixels themselves and generates samples based on them. This framework is equipped with an intervention mechanism which enables the model to generate samples from interventional distributions. We quantitatively and qualitatively assess the performance of CAN and empirically show that our model is able to generate both interventional and conditional samples without having access to the causal graph for the application of face generation on CelebA data.
中文速览
传统条件生成对抗网络（cGAN）在多标签图像生成时会独立采样各标签，完全忽略标签间的因果关系，导致生成"女性长胡须"这类不合常理的图像；而已有的CausalGAN虽然引入了因果推断框架，却要求事先给定因果图，在现实中难以满足。本文提出因果对抗网络（Causal Adversarial Network, CAN），其核心思路是把结构因果模型（Structural Causal Model, SCM）直接嵌入GAN的生成器，通过对抗训练从数据中自动学习标签之间的因果结构，无需人工提供因果图。CAN由两个模块构成：标签生成网络（LGN）负责学习并采样标签间的因果分布，条件图像生成网络（CIGN）则根据采样得到的标签学习标签与像素之间的关系并生成图像；此外，论文还设计了一种干预机制，通过掩码向量控制是否对某个节点施加外部干预，从而实现对干预分布（interventional distribution）的采样。在CelebA人脸数据集上的实验表明，CAN能够在不依赖预设因果图的情况下，同时生成高质量的条件样本和干预样本，这对于回答"如果这个人留了胡须会怎样"之类的反事实问题具有重要意义。
原文 arXiv:2008.11376;中英对照 + 大白话阅读 https://aha.fim.ai/paper/2008.11376v2