Dataset Condensation with Color Compensation
Dataset condensation faces inherent trade-offs between performance and fidelity. Current methods struggle with inefficiency (image-level selection) or semantic distortion (pixel-level optimization). We identify color’s dual role as an information carrier and semantic unit as critical. To address this, we propose DC3, which enhances color diversity via latent diffusion models after calibrated image selection. Experiments show DC3 outperforms SOTA methods across benchmarks. Notably, DC3 enables fine-tuning pre-trained diffusion models with condensed datasets without degradation (validated by FID scores), marking the first exploration of this capability.

FMCHS: Advancing Traditional Chinese Medicine Herb Recommendation with Fusion of Multiscale Correlations of Herbs and Symptoms
Traditional Chinese medicine (TCM) exhibits remarkable therapeutic efficacy in disease treatment and healthcare through personalized herb prescriptions. However, current herb recommendation models inadequately capture the multiscale relations between herbs and clinical symptoms, particularly neglecting latent correlations at the chemicalmolecular scale. To address these limitations, we propose the Fusion of Multiscale Correlations of Herbs and Symptoms (FMCHS), an innovative framework that synergistically integrates molecular-scale chemical characteristics of herbs with clinical symptoms. The framework employs multi-relational graph transformer layers to generate enriched embeddings that preserve both structural and semantic features within herbs and symptoms. Through systematic incorporation of herb chemical profiles into node embeddings and implementation of attention-based feature fusion, FMCHS effectively utilizes multiscale correlations.

Multimodal Classification of Temporally Relevant Images Using Large Pretrained Models
Temporally relevant images are visual representations that capture and convey information directly tied to specific points or periods in time. These images may exhibit gradual transformations in appearance, texture, or other visual cues that are challenging for traditional classification algorithms to recognize. To more accurately capture the features of temporally relevant images, we introduce spectral data corresponding to these images and propose a multimodal image classification method. The proposed approach leverages the expert knowledge of the large language model (i.e. Claude-3.5, GPT-4o) to guide a grounded vision language model and a segment anything model for zero-shot object detection and image segmentation.

VEX Robotics Competition
Led the school’s VEX Robotics club, responsible for programming and debugging of the robotic systems.
Attained Gold Awards at China Zone Selections, the Asia Championships, Asia Open and the World Championships in the United States during the 2018 season.