Research On Adversarial Learning For Data Distribution Matching

Generative Adversarial Networks(GANs)have become a popular research direction in the fields of computer vision,natural language processing and speech processing.However,most existing GAN methods have some challenges in the data distribution matching problem(such as data generation and multiple marginal matching).(1)Data generation aims at generating data from some prior distribution such that the generated data distribution matches the real data distribution.Unfortunately,such prior distribution is often independent of real data and thus may lose semantic information.In practice,the semantic information might be represented by some latent distribution learned from data,which,however,is hard to be used for sampling in GANs.(2)Multiple marginal matching problem aims at learning mappings to match a source domain distribution to multiple target domain distributions.Existing methods are intractable to measure the multi-marginal distance among different domains and exploit domain correlations to match the target domain distributions.For the data generation problem,this thesis proposes generative adversarial networks(LCC-GANs)based on local coordinate coding to match the real data distribution.Based on the data manifold hypothesis,this thesis applies an Auto Encoder to learn the data manifold to obtain the semantic information of real data.Then,this thesis proposes a sampling method to construct a local coordinate system on the manifold,which can effectively help LCC-GANs to sample meaningful points to generate new data.In theory,this thesis defines and derives the generalization bound of LCC-GANs,and proves that a small dimensional input is sufficient to achieve good generalization performance.Finally,extensive experiments demonstrate the effectiveness of the proposed method on real-world datasets.Moreover,the quality and diversity of the generated data are significantly higher than most existing GAN methods.For the multiple marginal matching problem,this thesis proposes multi-marginal Wasserstein distance-based generative adversarial networks(MWGANs)from a source domain distribution matching to multiple target domain distributions.Based on the optimal transport theory,this thesis first proposes a new dual formulation to measure and optimize the distribution distance among multiple domains.Relying on the dual formulation,this thesis proposes a new objective function for generative adversarial networks,which can exploit cross-domain correlations to match the target domain distributions by satisfying inner-domain and interdomain constraints.In theory,this thesis defines and analyzes the generalization performance of MWGANs on the multiple marginal matching problem,and derives its generalization bound.Finally,extensive experiments demonstrate the effectiveness of MWGANs on balanced and imbalanced translation tasks on real-world datasets.