中国科学院自动化研究所机构知识库

近十年内随着算力和数据爆发式增长，人工智能领域掀起了深度学习的研究热潮。生成对抗网络模型(Generative Adversarial Nets : GANs)作为一类深度生成模型成为一个研究热点。在过去的几年里，生成对抗网络模型在数据生成、图像修复、图像分割、图像超分辨率等领域获得了巨大成功，凸显了生成对抗网络模型的巨大潜力。生成对抗网络模型是从有监督学习模型到无监督学习模型的一次良好探索，是实现通用人工智能过程中的又一重大成果。因此对生成对抗网络模型的研究具有重要的理论意义和应用价值。
生成对抗网络模型存在两个基本问题：模式坍塌和训练不稳定。模式坍塌问题表现为生成对抗网络模型倾向于生成单一的、相似的数据，生成数据多样性匮乏。训练不稳定问题表现为网络性能指标在训练过程中不断震荡，难以收敛。这两个基本问题制约了生成对抗网络模型的发展和应用。本课题围绕模式坍塌和训练不稳定两个问题展开研究。针对模式坍塌问题，我们使用生成数据的熵来反映样本多样性，提出利用Stein Variational Gradient Descent (SVGD)算法实现生成数据熵惩罚的Stein Wasserstein GAN (SW-GAN)模型。针对训练不稳定问题，我们基于生成对抗网络模型中概率距离度量不合理优化的分析，提出利用Quantile Regression (QR)算法实现生成对抗网络模型中Wasserstein-1距离最小化的Quantile Regression GAN (QR-GAN)模型。本课题以生成对抗网络的代表工作Wasserstein GAN为基线模型。通过在MNIST、CIFAR-10数据集上的多样性分析实验，验证了SW-GAN模型增加样本多样性的有效性。通过在CIFAR-10、STL-10、LSUN-Tower等通用数据集的上分析实验，展示了QR-GAN模型在生成数据质量以及多种优化算法间性能稳定两个方面的优越性。

本课题从GANs模型的两个基本问题出发展开研究，研究成果主要有两点：其一，针对生成对抗网络模型的模式坍塌问题，提出了利用SVGD算法实现生成数据熵惩罚的改进方案。解决了生成对抗网络模型中熵无法显式求解而导致熵惩罚无法实现的困境，提高了生成样本的多样性。其二，针对训练不稳定问题，将适用于一维随机变量的QR算法扩展至生成对抗网络模型中，提高了生成数据的质量和模型在多种优化算法下的稳定性。为生成对抗网络模型中Wasserstein-1距离的实现提供新的求解思路。

In the past decade, with the explosion of computing power and data, the field of artificial intelligence has witnessed a frenzy of deep learning. Generative Adversarial Nets (GANs) as a deep generative model has received more and more attention. In the past few years, Generative Adversarial Nets model has achieved great success in data generation, image restoration, image segmentation, image super-resolution and other fields, which highlights the great potential of the Generative Adversarial Nets model. Generative Adversarial Nets model is a good exploration from supervised learning model to unsupervised learning model, and is another important achievement in the realization of general artificial intelligence. Therefore, the research on Generative Adversarial Nets model has important theoretical significance and application value.

There are two basic problems of Generative Adversarial Nets model: mode collapsing and training instability. Mode collapsing means that the generator tends to produce only a single sample or a small family of very similar samples,  and the generated data is lack of diversity. Training instability involves the generator and discriminator oscillating rather than converging on a fixed point in training. These two basic problems restrict the development and application of Generative Adversarial Nets model. So we focus on mode collapsing and training instability problem.  In order to alleviate the problem of mode collapsing, we propose Stein Wasserstein GAN (SW-GAN) model. The entropy of generated data is used to reflect the diversity of samples.  We add a generated distribution entropy term to the objective function of generator and maximize the entropy to increase the diversity of generated data. And then Stein Variational Gradient Descent (SVGD) algorithm is used for optimization. Aiming at the training instability problem, based on the analysis that the training instability problem originates from the unreasonable objective function, we propose  Quantile Regression GAN (QR-GAN) model. It  minimizes the Wasserstein-1 distance between true data and generated data with the help of Quantile Regression (QR) algorithm. Wasserstein GAN, which is the representative work of GANs model, is selected as the baseline model. Through the diversity analysis experiments on MNIST and CIFAR-10, the effectiveness of SW-GAN model for mode collapsing is verified. Experiments on CIFAR-10, STL-10, LSUN-Tower show the advantages of QR-GAN model in terms of data quality and stable performance among various optimization algorithms.

Based on the research of two basic problems of GANs model, there are two main achievements in this paper: firstly, aiming at the mode collapsing problem, we propose an improved method to realize entropy penalty of generated data by the help of SVGD algorithm. It solves the dilemma that entropy cannot be calculated in Generative Adversarial Nets model and resulting in the failure of entropy penalty. And it improves the diversity of generated data. Secondly, aiming at the problem of training instability, we extend QR algorithm for one-dimensional random variables to Generative Adversarial Nets model, which improves the quality of generated data and the stability of the model under multiple optimization algorithms. It provides a new way to realize Wasserstein-1 distance in Generative Adversarial Nets model.