04/12
2023
Seminar
- Title题目 Investigating the Inference Suboptimality of the Conditional Generative Model and How to Fix It
- Speaker报告人 王琦 (National University of Defense Technology)
- Date日期 2023年4月12日 10:00
- Venue地点 南楼 6420
Abstract摘要
The recent few years have witnessed the great potential of conditional generative models. Some works like GPT4 and Stable Diffusion exhibit incredible performance in Artificial Intelligence Generated Content (AIGC). This talk will focus on the fundamental issues of conditional generative models and take the neural process model (Garnelo et al., 2018) as an example to investigate the inference suboptimality of variational inference. To close the inference gap, I overview the available optimization objectives and construct the surrogate objective inspired by the variational expectation maximization. The resulting model, referred to as the Self-normalized Importance weighted Neural Process (SI-NP), can learn a more accurate functional prior and has an improvement guarantee concerning the target log-likelihood. Experimental results show the competitive performance of SI-NP over other objectives, guiding the design of inference algorithms for conditional generative models.
Reference: Qi Wang, Marco Federici, & Herke van Hoof, Bridge the Inference Gaps of Neural Processes via Expectation Maximization, ICLR2023.
个人简介：Qi Wang is currently an assistant professor in statistics and operation research at the National University of Defense Technology. Before joining NUDT, he obtained a Ph.D. in machine learning under the supervision of Prof. Max Welling and Dr. Herke van Hoof. During that time (2019.06-2022.09), his research interest was in generative modeling and reinforcement learning. He published several papers in ICML, NeurIPS, and ICLR and developed a couple of conditional generative models, meta-reinforcement learning models, and inference algorithms. He received an ICML2022 travel grant and a NeurIPS2022 scholar award.
邀请人： 张潘
附件下载: