|
Welcome to my website~ My name is Shengkun Tang. You can call me Bryson for short. Currently, I am a first-year PhD student of Machine Learning in MBZUAI, under the supervision of Prof. Zhiqiang Shen. During my gap year, I had a wonderful time as an research assistant in DASLab in ISTA , working with Prof. Dan Alistarh. Besides, I had close collaboration with Prof. Dongkuan Xu (NCSU) and Dr. Yaqing Wang (Google DeepMind), working on efficent multi-modal models. I finished B.E. in Remote Sensing at Wuhan University , under the supervision of Prof. Jian Yao and Prof. Xin Su. Email / Google Scholar / Github Last updated: Feb. 4th 2025 |
|
 |
Cong Zeng*, Shengkun Tang*, Xianjun Yang, Yuanzhou Chen, Yiyou Sun, Yao Li, Haifeng Chen, Wei Cheng, Dongkuan Xu [NeurIPS 2024] The Thirty-eighth Annual Conference on Neural Information Processing Systems arXiv / code We propose a simple but quite effective method to improve the performance of black-box LLM detection. DALD collects a small-size data from target model and train the surrogate model to align the distribution of surrogate model and target model. |
 |
Shengkun Tang, Yaqing Wang, Caiwen Ding, Yi Liang, Yao Li, Dongkuan Xu [ECCV 2024] European Conference on Computer Vision arXiv / code We propose a uncertainty estimation module (UEM) to decide the exiting point during diffusion model inference at each timestep. Moreover, we propose an uncertainty-aware layer-wise loss to recover the performance for early-exited model. |
 |
Shengkun Tang, Yaqing Wang, Zhenglun Kong, Tianchi Zhang, Yao Li, Caiwen Ding, Yanzhi Wang, Yi Liang, Dongkuan Xu [CVPR 2023] The IEEE/CVF Conference on Computer Vision and Pattern Recognition arXiv / code We propose a novel early exiting strategy based on cascading input similarity with valid assumptions on saturation states in visual-language models, a pioneering exploration of extending early exiting selection to encoders and decoders of sequence-to-sequence architectures. |
 |
Puyuan Yi*, Shengkun Tang*, Jian Yao Preprint, 2021 arXiv / code We propose a Dynamic Depth Range Network (DDR-Net) to determine the depth range hypotheses dynamically by applying a range estimation module (REM) to learn the uncertainties of range hypotheses in the former stages. |
|
Haonan Guo, Xin Su, Shengkun Tang, Bo Du, Liangpei Zhang IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2021 We propose a novel deep-supervision convolutional neural network (denoted as DS-Net) for extracting building footprints from high-resolution remote sensing images. |
|
Vision Algorithm Intern Researcher Project: SenseRobot Chess Robotic, working with Ruodai Li |
|
|
System Developer Project: Large-Scale Self-Driving System Development |
|
Ranking: 7/2305 (teams) The task of Baidu Astar 2020 is traffic signs and surveillance cameras detection and matching. I was in charge of detection task. I solved the problems of data imbalance by using my own data argumentation strategy and detect surveil- lance cameras more accurately. We got into the final and rank 7 out of 2305 teams. |
|
|
|
|
This template comes from source code, thanks for his fantastic website templates.
|