Chao Chen

Contact

• Office: Computer Science Building, 2313C
• Tel: +1-631-632-2593
• Email: chao.chen.1 (@) stonybrook.edu

Research Interests

I explore geometric and topological properties of data. These global and robust information can provide insight for modern data and image analytics. My research draws from the following different domains.

• Biomedical imaging informatics: topology-informed image segmentation, generation and analysis.

• Machine learning: robustness of deep neural networks, label noise, graph neural networks.

• Topological data analysis: persistent homology, computation and learning with topological features.

Past Experience

• Assistant Professor of Computer Science, City University of New York (CUNY), United States

• Postdoc at Rutgers University, United States (advised by Prof. Dimitris Metaxas)

• Postdoc at Institute of Science and Technology, Austria (advised by Prof. Herbert Edelsbrunner)

• Ph.D. at Rensselaer Polytechnic Institute, United States (advised by Prof. Daniel Freedman)

• B.S. at Peking University, China

Recent Services

• Associate Editor, Pattern Recognition

• Area Chair, CVPR 2022

• Area Chair, NeurIPS 2021, 2022

• Area Chair, MICCAI 2021, 2020

• Senior PC, AAAI 2022, 2021

News and Annoucement

• New!! One paper accepted by ICML 2022!

  We develop a cycle-centric graph neural network which learns representations of cycles instead of nodes. The method is applied to knowledge graph completion task.

• New!! One paper accepted by NAACL 2022!

  We provide an insightful analysis of backdoor-attacked BERT models focusing on the attention focus drifting phenomenon. This provides an explanation of how backdoors are injected and inspires better backdoor detection algorithms.

• New!! One paper accepted by CVPR 2022!

  We improve predictions on snapshot medical images using temporal sequences. We design a model that use temporal representation to re-calibrate snapshot representation.

• New!! Just won the NSF CAREER award.

  Project: Topology-Driven Learning for Biomedical Imaging Informatics

  This 5 year project will support us in developing topology-informed image analysis methods for different tasks such as segmentation, generation and analysis. The key is to combine theory from topological data analysis with deep learning techniques to develop trasparent, trustworthy and powerful tools. These tools will be validated in various biomedical problems.

• New!! One paper accepted by SoCG 2022!

  We develop a GPU algorithm to speed up the computation Euler Characteristics Curve for large scale imaging data.

• New!! One paper accepted by ICLR 2022!

  We use a topological prior to better reconstruct triggers of the input. This helps us identify backdoor attacked neural network models.

• New!! One paper accepted by AISTATS 2022!

  We investigate adversarial robustness from a manifold's perspective.

• New!! One paper accepted by NeurIPS 2021!

  We use the tool of persistent homology to investigate the behavior of neural networks under backdoor attacks.

• New!! One paper accepted by ICCV 2021!

  We model spatial context for cell detection and classification in histopathology images. We use the k-function in spatial statistics and incorporate into the feature representation on our deep neural network.

• New!! One paper accepted by ICML 2021!

  We use topological features for link prediction with graph neural networks! Another contribution of the paper is a quadratic algorithm for the computation of extended persistent homology feature with graphs. It can be applied to any graph learning methods.

• New!! One paper accepted by IPMI 2021!

  We learn topological biomarkers for MRI images of breast cancer patients.

• New!! Two papers accepted by ICLR 2021 as spotlight presentations!

  We use discrete Morse theory to improve the topology-preserving segmentation problem.

  We also introduce a novel noise model and a theoretically-guaranteed algorithm to train robust model against such noise.

• New!! One paper accepted by AAAI 2021!

  We use topological information for crowd counting.

• New!! Two papers accepted by NeurIPS 2020!

  We train robust models by filtering out label noise using topological information. We also proposed a new variational formulation for the instance segmentation task.