cv

Visiting Student

ENCODE Lab, Westlake University

Advised by Prof. Huan Wang. Focused on iOS mobile model deployment and visual autoregressive model compression.

• Conducted research on iOS model deployment
• Developed compression methods for visual autoregressive models

Research Assistant

Zhejiang University

Industrial VR Glasses Backend Image Recognition Algorithm Development for Tower Group.

• Built industrial electrical cabinet component dataset for detecting 17 types of electrical components
• Implemented electrical cabinet label text recognition using Paddle OCR to verify device label correctness
• Deployed algorithms on server with proficiency in Linux and Docker workflows

Master

Zhejiang University

Control Engineering

Bachelor

North University of China

Rail Transit Signal and Control

OBS-Diff: Accurate Pruning For Diffusion Models in One-Shot

ICLR (Under Review)

Fourth author. Proposed an OBS-based training-free multi-granularity pruning method for diffusion models with timestep-aware Hessian construction and logarithmically decreasing weights to mitigate error accumulation.

EVAR: Edge Visual Autoregressive Models via Principled Pruning

CVPR (Under Review)

First author. Proposed a principled OBS-based structured pruning method for visual autoregressive models with progressive scale-aware distillation to address gradient imbalance. Achieved 1.8× speedup with only 10% quality loss on edge devices for single-image generation.

Network Binarization via Contrastive Deep Supervision

NCAA Conference

First author. Proposed BNN-CDS, a binary network optimization method based on deep supervision and contrastive learning fusion, with a suitable scaling factor strategy.

An Effective Information Theoretic Framework for Channel Pruning

IEEE Transactions on Neural Networks and Learning Systems (TNNLS)

Second author. Proposed an information-theoretic framework using entropy and rank fusion for layer-wise pruning rates, with Shapley value-based contribution evaluation as the intra-layer pruning criterion.

EVAR: Edge Visual Autoregressive Model Compression

Developed principled pruning methods for visual autoregressive models targeting edge device deployment. Introduced progressive scale-aware distillation to handle gradient imbalance during autoregressive fine-tuning.

• 1.8× speedup on edge devices
• 10% quality loss in single-image generation
• OBS-based structured pruning

Industrial Electrical Cabinet Recognition System

Developed an image recognition system for industrial VR glasses to detect and verify electrical cabinet components and labels.

• 17 component types detection
• OCR-based label verification
• Server deployment with Docker