Hao Zhang

I am now working in Xidian University, as a professor. From 08/01/2020 to 07/31/2022, I was a postdoctoral researcher in the Department of Population Health Sciences at Weill Cornell Medicine, advised by Professor Fei Wang. From 08/01/2019 to 07/31/2020, I was a postdoctoral researcher in the Department of Electrical and Computer Engineering at Duke University, advised by Professor Lawrence Carin.

I received my B.Sc. and Ph.D. from Xidian University in 2012 and 2019, respectively. My Ph.D. advisor was Professor Bo Chen. I was also advised by Professor Mingyuan Zhou, an associate professor from the University of Texas at Austin.

My research lies at the intersection of Bayesian statistics and machine learning. I am interested in statistical methods, hierarchical models, statistical inference for big data, and deep learning. I did some researches about applying machine learning methods on Electronic Health Records (EHRs) analysis when I worked in WCM. I currently focus on advancing both statistical inference with deep learning, deep learning with probabilistic methods and their applications to natural language processing, multi-modal learning, and radar signal processing.

Research Highlights

[Aug, 2024] Our paper "HICEScore: A Hierarchical Metric for Image Captioning Evaluation" is accepted by ACM MM 2024. In this work, we propose a novel reference-free metric for the evaluation of image-text similarity, dubbed Hierarchical Image Captioning Evaluation Score (HICE-S). We will release the code and paper can be found in Arxiv.

[May, 2024] I passed the defense of the professorship.

[February, 2024] Our paper "MeaCap: Memory-Augmented Zero-shot Image Captioning" is accepted by CVPR2024. This work develops a new zero-shot image captioning method developed from our CVPR2023 work, ConZIC. Compared with ConZIC, MeaCap can generate concept-centered captions that keep high consistency with the image with fewer hallucinations and more world-knowledge. We will release the code and paper can be found in Arxiv.

[December, 2023] Our paper "High-throughput target trial emulation for Alzheimer’s disease drug repurposing with real-world data" is accepted by Nature Communications. This work develops a new machine learning based target trial emulation methods. We also use this method for AD drug repurpose. We release the code and paper can be found in Nature Communications

[September, 2023] Our paper "Single-cell multi-omic topic embedding reveals cell-type-specific and COVID-19 severity-related immune signatures" is accepted by Cell Report Methods. This work develops a method of embedding topic model for multi-omics data anlaysis, including cell clustering and cross-domain imputation. We also use this method to reveals cell-type-specific and COVID-19 severity-related immune signatures. We release the code, and reported by a Chinese report. See 岚翰生命科学

[Augest, 2023] Our paper "Contrastive Attraction and Contrastive Repulsion for Representation Learning" with my friends is accepted by Transactions on Machine Learning Research. This work introduces a doubly contrastive-learning strategy that contrasts positive samples and negative ones within themselves separately, to improve the performance of contrastive learning and enhance its robustness on various datasets. We release the code

[July, 2023] Our paper "Regularized Data Programming with Automated Bayesian Prior Selection" with my Cornell students and professors is accepted by ICML2023 workshop. This work introduces a Bayesian extension of classical DP that mitigates failures of unsupervised learning by augmenting the DP objective with regularization terms.

[March, 2023] We released the code of ConZIC accepted by CVPR2023, and the paper is online on Arxiv.

[March, 2023] I passed the defense of the associate professor.

[March, 2023] Our paper "ConZIC: Controllable Zero-shot Image Captioning by Sampling-Based Polishing" with my students from Xidian University is accepted by CVPR2023. In this work, we propose a framework for Controllable Zero-shot Image Captioning, named ConZIC. The core of ConZIC is a novel sampling-based non-autoregressive language model named Gibbs-BERT, which can generate and continuously polish every word. We will release the code.

[Jan, 2023] Our paper "Data-driven identification of post-acute SARS-CoV-2 infection subphenotypes" is reported by 自然界, a Chinese report. See 自然界

[Jan, 2023] Our paper "Data-driven identification of post-acute SARS-CoV-2 infection subphenotypes" has been selected as the Research Briefing of Nature Medicine. See Nature Medicine

[Dec, 2022] Our paper "Data-driven identification of post-acute SARS-CoV-2 infection subphenotypes" with my collaborator from Weill Cornell Medicine is published in Nature Medicine. In this work, we use the topic modeling to analysis the large-scale EHR dataset and we identified four reproducible PASC subphenotypes, dominated by cardiac and renal; respiratory, sleep and anxiety; musculoskeletal and nervous system; and digestive and respiratory system sequelae.

[Sep, 2022] I joined the group of Professor Hongwei Liu and Bo Chen at Xidian University, as a teacher.

[Sep, 2022] Our paper "A Variational Edge Partition Model for Supervised Graph Representation Learning" with Yilin He, Chaojie Wang, Bo Chen, and Mingyuan Zhou will be published in NeurIPS 2022. In this work, we propose a graph generative process to model how the observed edges are generated by aggregating the node interactions over a set of overlapping node communities, each of which contributes to the edges via a logical OR mechanism. Extensive evaluations on real-world graph datasets have verified the effectiveness of the proposed method in learning discriminative representations for both node-level and graph-level classification tasks.

[Sep, 2021] Our paper "Multiscale Visual-Attribute Co-Attention for Zero-Shot Image recognition" with Long Tian, Zhengjue Wang, Yishi Xu, Pengyu Cheng, Ke Bai, and Bo Chen will be published in IEEE TNNLS. In this work, we propose a multi-scale visual-attribute co-attention (mVACA) model, considering both visual-semantic alignment and visual discrimination at multiple scales.

[Sep, 2021] Our paper "Learning Hierarchical Document Graphs From Multilevel Sentence Relations" with Chaojie Wang, Zhengjue Wang, Zhibin Duan, Bo Chen, Mingyuan Zhou, Ricardo Henao and Lawrence Carin will be published in IEEE TNNLS. This work is about learning a set of learnable hierarchical graphs via exploring multilevel sentece relations, assisted by a hierarchical probabilistic topic model.

[May, 2021] Our paper "EnsLM: Ensemble Language Model for Data Diversity by Semantic Clustering" with Zhibin Duan, Chaojie Wang, Zhengjue Wang, Bo Chen, and Mingyuan Zhou will be published in ACL 2021. This work is about by semantic clustering, building an ensemble language model to alleviate the heterogeneous charateristics of data unsupervisedly. Our idea can be used for language generation (decoder) and understanding (encoder and autoencoder).

[May, 2021] Our paper "Multimodal Weibull Variational Autoencoder for Jointly Modeling Image-Text Data" with Chaojie Wang, Bo Chen, Sucheng Xiao, Zhengjue Wang, Penghui Wang, Ning Han, and Mingyuan Zhou will be published in IEEE Trans. on Cybernetics. This work is about building an interpretable image-text modalities probability autoencoder.

[March, 2021] Our paper "MetaSCI: Scalable and Adaptive Reconstruction for Video Compressive Sensing" with Zhengjue Wang, Ziheng Cheng, Bo Chen, and Xin Yuan will be published in CVPR2021. This work is about developing a video snapshot compressive imaging (SCI) reconstruction method based on meta learning for large-scale videos and for fast adpation to a new mask.

[March, 2021] Our paper "Memory-Efficient Network for Large-scale Video Compressive Sensing" with Ziheng Cheng, Bo Chen, Guanliang Liu, Ruiying Lu, Zhengjue Wang, and Xin Yuan will be published in CVPR2021. This work is about developing a memory-efficient video snapshot compressive imaging (SCI) reconstruction method for large-scale videos.

[December, 2020] Our paper "Max-Margin Deep Diverse Latent Dirichlet Allocatioin with Continual Learning" with Wenchao Chen, Bo Chen, Yingqi Liu, Qianru Zhao, and Long Tian will be published in IEEE Transactions on Cybernetics. This work is about developing a deep deverse latent Dirichlet alloation with continual learning to solve the problem of high-layer topics in DLDA without diversity.

[September, 2020] Our paper "Bidirectional Convolutional Poisson Gamma Dynamical Systems" with Wenchao Chen, Chaojie Wang, Yicheng Liu, Bo Chen, and Mingyuan Zhou will be published in NeurIPS2020. This work is about building a bidirectional convolutional topic modeling for document classifcation and exploring the relations among sentences in one document.

[September, 2020] Our paper "Deep Relational Topic Modeling via Graph Poisson Gamma Belief Network" with Chaojie Wang, Zhengjue Wang, Dongsheng Wang, Bo Chen, and Mingyuan Zhou will be published in NeurIPS2020. This work is about using hierarical topic model to explore the graph data for node clustering, node classification and node-relation prediction.

[September, 2020] Our paper "Friendly Topic Assistant for Transformer Based Abstractive Summarization" with Zhengjue Wang, Zhibin Duan, Chaojie Wang, Long Tian, Bo Chen, and Mingyuan Zhou will be published in the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP2020). This work is about using topic model to help Transformer based language model for document abstractive summarization.

[September, 2020] Our paper "Unsupervised Hyperspectral and Multispectral Image Fusion Based on Nonlinear Variational Probabilistic Generative Model" with Zhengjue Wang, Bo Chen, and Hongwei Liu will be published in IEEE Transactions on Neural Network and Learning System. This work is about building an probabilistic autoencoding framework for hyperspectral and multispectral image fusion.

[August, 2020] I joined the group of Professor Fei Wang at Weill Cornell Medicine, as a posdoctoral researcher.

[July, 2020] Our paper "BIRNAT: Bidirectional Recurrent Neural Networks with Adversarial Training for Video Snapshot Compressive Imaging" with Ziheng Cheng, Ruiying Lu, Zhengjue Wang, Bo Chen, Ziyi Meng, and Xin Yuan will be presented in ECCV2020. This work is about the problem of reconstruction of video snapshot compressive imaging (SCI) and we build a bidirectional RNN for this task. Our work achieves the SOTA performance on this task among existing deep learning methods.
[June, 2020] Our paper "FusionNet: An Unsupervised Convolutionl Variational Network for Hyperspectral and Multispectral Image Fusion" with Zhengjue Wang, Bo Chen, Ruiying Lu, Hongwei Liu, and Varshney will be published in IEEE Transactions on Image Processing. This work is about building an convolutional variational autoencoder for hyperspectral and multispectral image fusion.

[June, 2020] Our paper "Students Need More Attention: BERT-based Attention Model for Small Data with Application to Automatic Patient Message Triage" with Shijing Si, Rui Wang, Jedrek Wosik, David Dov, Guoyin Wang, Ricardo Henao, and Lawrence Carin will be presented in MLHC2020. This work is to build a model based on BERT for patient portal message triage that classifies the urgency of a message into three categories: non-urgent, medium and urgent.

[June, 2020] Our paper "Autoencoding Topic Model with Scalable Hybrid Bayesian Inference" with Bo Chen, Yulai Cong, Dandan Guo, Hongwei Liu, and Mingyuan Zhou will be published in IEEE Transactions on Pattern Analysis and Machine Intelligence. This work is about building an autoneconding hierarchical topic model for document analysis.

[Janurary, 2020] Our paper "Learning Dynamic Hierarchical Topic Graph with Graph Convolutional Network for Document Classication" with Zhengjue Wang, Chaojie Wang, Zhibin Duan, Bo Chen, and Mingyuan Zhou will be presented in AISTATS2020. This work is about building a dynamic document graph with the help of a hierarhcical topic model for document classification. Hope to see you in Palermo, Sicily, Italy, in June 2020.

[December, 2019] Our paper "Variational Hetero-Encoder Randomized GANs for Joint Image-Text Modeling" with Long Tian, Bo Chen, Zhengjue Wang, and Mingyuan Zhou will be presented as a poster in ICLR2020. This work is about a novel Bayesian deep learning framework that captures and relates hierarchical semantic and visual concepts, performing well on a variety of image and text modeling and generation tasks. Hope to see you in Addis Ababa, Ethiopia in April, 2020.

[August, 2019] I joined the group of Professor Lawrence Carin at Duke University, as a posdoctoral researcher.

[June, 2019] I recieved my Ph.D. from Xidian University advised by Professor Bo Chen, Xi'an, China. Thanks for Professor Bo Chen and Professor Mingyuan Zhou from UT-Austin.

Honors and Awards

[2013] Chinese National Scholarship for master.

[2015] Chinese National Scholarship for Ph.D. student.