|
Ph.D. student from the IIML Lab at the University College London, Department of Electronic and Electrical Engineering, advised by Prof. Miguel Rodrigues. Research interests include AI reliability, large language models, vision-language models, multimodal learning, and machine learning. |
|
|
Topics: AI Reliability; Large Language Models; Machine Learning; Multimodal Learning. Topics: Compressed Sensing; Machine Learning; Imaging Systems. Topics: Pedestrian Tracking; Kalman Filtering; Image Processing. |
|
|
|
Chen Feng†, Minghe Shen†, Ananth Balashankar, Carsten Gerner-Beuerle, Miguel R. D. Rodrigues † Equal contribution. International Conference on Learning Representations (ICLR), 2026 [Paper] We introduce a “Noisy but Valid” testing framework that uses a small human-labeled calibration set to estimate judge error rates and applies a variance-corrected threshold to large judge-labeled data. Our method guarantees finite-sample Type-I error control despite calibration uncertainty. We provide theoretical conditions for improved power, empirical validation on benchmark datasets, and quantify the gap to an ideal “Oracle” judge—offering the first systematic treatment of certification with imperfect LLM evaluators. |
|
Minghe Shen, Hongping Gan, Chunyan Ma, Chao Ning, Hongqi Li, and Feng Liu IEEE Transactions on Cybernetics (TCYB), 2024 [Paper] [Code] We propose MTC-CSNet, a hybrid dual-path framework combining a lightweight ConvNet branch for local feature extraction and a Transformer branch for global correlation modeling. A bridging unit adaptively fuses both features for high-quality reconstruction. |
|
Hongping Gan, Minghe Shen, Yi Hua , Chunyan Ma, and Tao Zhang IEEE Transactions on Computational Imaging (TCI), 2023 [Paper] [Code] We propose TCS-Net, a hierarchical Transformer-based framework for image CS using a patch-to-pixel strategy. It includes a learnable acquisition module and a multi-stage reconstruction module that progressively decodes measurements from patch-level structure to pixel-level details. |
|
Chao Ning, Hongping Gan, Minghe Shen, Tao Zhang Engineering Applications of Artificial Intelligence (EAAI), 2023 [Paper] [Code] We propose learning based padding that predicts border values from feature map context using lightweight convolution (LPC) and attention (LPA) modules, providing a generic plug and play replacement for zero padding that consistently improves accuracy on image classification and semantic segmentation across diverse backbones. |
|
Minghe Shen, Hongping Gan, Chao Ning, Yi Hua, and Tao Zhang IEEE Transactions on Image Processing (TIP), 2023 [Paper] [Code] We propose TransCS, a Transformer-based hybrid framework for high-quality image CS. It uses a trainable sensing matrix for data-driven sampling and a customized ISTA-based Transformer backbone to model global dependencies during iterative reconstruction. An auxiliary CNN captures local image features. |
|
|
|
|
|
|