Dandan Guo

I am currently a Postdoc at The Chinese University of Hong Kong, Shenzhen (2020.12-2022.12). I am advised by Professor Hongyuan Zha, a X.Q. Deng Presidential Chair Professor of the Chinese University of Hong Kong, Shenzhen and the Executive Dean of the School of Data Science. I received my Ph.D. from Xidian University in August,2020. My Ph.D. advisor was Professor Bo Chen. And I am also advised by Professor Mingyuan Zhou, an associate professor from the University of Texas at Austin.

My research lies at the intersection of statistical machine learning and its combinations with real-world applications. I am interested in probabilistic methods, deep generative models, representation learning and meta learning. These developed models and algorithms have been applied to time series modeling, text analysis, natural language processing, few-shot generation (classification), and automatic radar target recognition.

Research Highlights

[May, 2022] Our paper "Matching Visual Features to Hierarchical Semantic Topics for Image Paragraph Captioning" with Ruiying Lu, Bo Chen and Mingyuan Zhou will be published in IJCV 2022. In this paper, we develop a plug-and-play hierarchical-topic-guided image paragraph generation pipeline, which couples a visual extractor with a deep topic model to guide the learning of a language paragraph generation model.

[February, 2022] Our paper "Learning Prototype-oriented Set Representations for Meta-Learning" with Long Tian, Minghe Zhang, Mingyuan Zhou and Hongyuan Zha will be published in ICLR 2022. In this paper, we present a novel OT-based method to improve existing summary networks designed for set-structured input based on optimal transport, where a set is endowed with two distributions: one is the empirical distribution over the data points, and another is the distribution over the learnable global prototypes. Our plug and play framework has shown appealing properties that can be applied to many meta-learning tasks.

[February, 2022] Our paper "Representing Mixtures of Word Embeddings with Topic Embeddings" with Dongsheng Wang, He Zhao, Huangjie Zheng, Korawat Tanwisuth, Bo Chen and Mingyuan Zhou. " will be published in ICLR 2022. We introduce WeTe, a new topic modeling framework, which views the learning of a topic model as the process of minimizing the expected CT cost between those two sets over all documents.