Dequan Wang

Shanghai Jiao Tong University

Shanghai Artificial Intelligence Laboratory

Assistant Professor

Shanghai Jiao Tong University

Jan 2023 – Present Shanghai

Research Scientist

Shanghai Artificial Intelligence Laboratory

Jan 2023 – Present Shanghai

PhD in Computer Science

University of California, Berkeley

Aug 2016 – Dec 2022 Berkeley, CA

BSc in Computer Science

Fudan University

Sep 2012 – Jun 2016 Shanghai

Dequan Wang (王德泉) serves as an Assistant Professor at Shanghai Jiao Tong University and concurrently as a Research Scientist at Shanghai Artificial Intelligence Laboratory. His research interests are predominantly in computer vision and machine learning, specifically the application of generative AI within the realms of healthcare and life sciences.

Prior to these roles, Dequan Wang completed his Ph.D. in Computer Science at UC Berkeley under the guidance of Prof. Trevor Darrell. His undergraduate studies were undertaken at the School of Computer Science at Fudan University, where he worked closely with Prof. Yu-Gang Jiang, Prof. Zheng Zhang, and Prof. Xiangyang Xue.

Publications

MedFMC: A Real-world Dataset and Benchmark for Foundation Model Adaptation in Medical Image Classification

MedFMC, our novel dataset of 22,349 images from diverse clinical tasks, aims to bolster the use of pre-trained foundation models in medical image classification—a sector facing data scarcity. The application of MedFMC underscores the potential of these models and highlights the effectiveness of various baseline methods in terms of accuracy and cost-efficiency.

Dequan Wang, Xiaosong Wang, Lilong Wang, Mengzhang Li, Qian Da, Xiaoqiang Liu, Xiangyu Gao, Jun Shen, Junjun He, Tian Shen, Qi Duan, Jie Zhao, Kang Li, Yu Qiao, Shaoting Zhang

Text-guided Foundation Model Adaptation for Pathological Image Classification

Our method, Connect Image and Text Embeddings (CITE), enhances pathological image classification by integrating biomedical text knowledge. CITE proves its superiority, particularly in low-data scenarios, with the PatchGastric stomach tumor dataset. Therefore, CITE provides a novel approach to enhancing data-efficient pathological image analysis using in-domain text knowledge.

Yunkun Zhang, Jin Gao, Mu Zhou, Xiaosong Wang, Yu Qiao, Shaoting Zhang, Dequan Wang

Back to the Source: Diffusion-Driven Adaptation to Test-Time Corruption

Our method, Diffusion-Driven Adaptation (DDA), adapts test inputs to enhance model accuracy on shifted data. Utilizing a generative diffusion model with image guidance and classifier self-ensembling, DDA surpasses traditional model adaptation approaches in handling various types of data corruption, diverse data quantities, and dependencies, as confirmed by the ImageNet-C benchmark tests.

Jin Gao, Jialing Zhang, Xihui Liu, Trevor Darrell, Evan Shelhamer, Dequan Wang

GACT: Activation Compressed Training for Generic Network Architectures

GACT is an activation compression training (ACT) framework to support a broad range of machine learning tasks for generic neural network architectures with limited domain knowledge. By analyzing a linearized version of ACT’s approximate gradient, we prove the convergence of GACT without prior knowledge on operator type or model architecture.

Xiaoxuan Liu, Lianmin Zheng, Dequan Wang, Yukuo Cen, Weize Chen, Xu Han, Jianfei Chen, Zhiyuan Liu, Jie Tang, Joseph Gonzalez, Michael Mahoney, Alvin Cheung

Contrastive Test-time Adaptation

We introduce AdaContrast, a novel approach to test-time adaptation that leverages contrastive learning and an advanced online pseudo labeling scheme. AdaContrast incorporates a memory queue for label refinement, resulting in improved performance, enhanced memory efficiency, and better model calibration. It outperforms existing methods and demonstrates reduced sensitivity to hyperparameters.

Dian Chen, Dequan Wang, Trevor Darrell, Sayna Ebrahimi

On-target Adaptation

As for most adaptation methods, most of the parameter updates for the model representation and the classifier are derived from the source and not the target. However, target accuracy is the goal, so we argue for optimizing as much as possible on the target data.

Dequan Wang, Shaoteng Liu, Sayna Ebrahimi, Evan Shelhamer, Trevor Darrell

Fighting Gradients with Gradients: Dynamic Defenses against Adversarial Attacks

Dent improves adversarial/robust accuracy (%) by more than 30 percent (relative) against AutoAttack on CIFAR-10 while preserving natural/clean accuracy. The static defenses alter training, while dent alters testing, and so this separation of concerns makes dent compatible with many existing models and defenses.

Dequan Wang, An Ju, Evan Shelhamer, David Wagner, Trevor Darrell

ActNN: Reducing Training Memory Footprint via 2-Bit Activation Compressed Training

ActNN is a PyTorch library for memory-efficient training. It reduces the training memory footprint by compressing the saved activations. ActNN is implemented as a collection of memory-saving layers. These layers have an identical interface to their PyTorch counterparts.

Jianfei Chen, Lianmin Zheng, Zhewei Yao, Dequan Wang, Ion Stoica, Michael Mahoney, Joseph Gonzalez