Heat-Kernel Empowered Deep Generative Models
Deep generative model (DGM) has been an influential topic in deep learning with great success in various applications. Although there have been many works on generative adversarial networks, many of them do not take manifold information of the training data into consideration in training. In this talk, I will present our recent work to effectively achieve the goal by simultaneously learning an intrinsic heat kernel of the manifold. The heat kernel encodes extensive geometric information of a manifold in an implicit way. In our work, we propose a way to incorporate manifold information into kernel-based DGMs by substituting the kernel with the learned heat kernel in the DGM. Our experimental results on image synthesis demonstrate the superiority of the proposed method, obtaining better generation quality relative to strong baselines.
Key Takeaways: 1. A deep generative model to learning to incorporate manifold information from the training data. 2. Achieved by simultaneously learning the associated heat kernel of the manifold. 3. Well-justified theoretical guarantees and improved performance on several image generation tasks.
Changyou Chen is an Assistant Professor in the Department of Computer Science and Engineering at the University at Buffalo, State University of New York. His research interest includes Bayesian machine learning, deep learning and deep reinforcement learning. Previously, Dr. Chen was a Research Assistant Professor and a Postdoctoral Associate in the Department of Electrical and Computer Engineering, Duke University. He got his PhD from College of Engineering and Computer Science, the Australian National University.