Yu Li 李煜

AlphaAdam:Asynchronous Masked Optimization with Dynamic Alpha for Selective Updates
Da Chang, Yu Li, Ganzhao Yuan
2025.1
Preprint
[Paper]
Developed AlphaAdam, an LLM optimization framework using intra-layer parameter updates. Created parameter masks based on historical momentum and gradient consistency, paired with adaptive mask strength for efficient optimization and guaranteed convergence. Outperformed AdamW in convergence speed and efficiency across GPT-2, RoBERTa, and Llama-7B tasks.

Sensor Fusion for Robust and Efficient Autonomous Vehicle Perception
Yu Li, Junyao Wang
2024.6-2024.10
Submitted to ICCV 2025
[Code]

Our main work is to construct our VLM to improve the autonomous driving mode in abnormal weather conditions, focusing on improving the multi-modal fusion algorithm of Camera-Lidar branches. Hopefully it will be advanced to hardware deployment in the future.

Ultrasound image segmentation based on SAM and Transformer multi-level feature fusion
Yu Li, Jin Huang
2023.10-2024.5
Medical Physics

We leveraged the rich semantic segmentation information of SAM and applied its fine-grained attention capability to the feature extraction module of Transformer, achieving SOTA IoU scores in ultrasound image segmentation.

Windowed self-attention guided multi-scale feature stream alignment network for ultrasound image segmentation
Jin Huang, Xiaoxiao Li, Yu Li
2023.9-2024.5
Under Review, Biomedical Signal Processing and Control.

We explored the self-attention guidance mechanism of the Swin-Transformer structure in breast ultrasound images, used a multi-level encoder to deal with multi-dimensional features, and used spatial pyramid pooling and feature stream alignment to fuse shallow and deep information.

Diffusion model with discriminative priors: Self-supervised monocular depth estimation in endoscopy
Yu Li
2024.2-2024.6
Accept, International Journal of Computer Assisted Radiology and Surgery.

For endoscope medical scenarios, we use the diffusion model for depth estimation. We build a teacher model, set knowledge distillation, optical appearance and ddim losses, and introduce the teacher's discriminative prior, which significantly enhances the accuracy and confidence of the results.