Xuefeng Jiang (蒋雪枫)

I am a forth-year PhD candidate at Institute of Computing Technology, Chinese Academy of Sciences in Beijing, China. Before that, I received my Bachelor degree from Beijing University of Posts and Telecommunications in the summer of 2021.

I am advised by Prof. Min Liu, associate Prof. Yuwei Wang and associate Prof. Sheng Sun, in the Mobile Computing Team. My research interest mainly includes distributed ML system, data quality, cyber security, AutoDrive and various deep learning applications. I am always open to discussions, please feel free to contact me!

I am on the job market for 2025, and I'm currently seeking for suitable working and intern chances. Here's my resume. Welcome to contact me!

Email  /  Google Scholar  /  Github  /  OpenReview  /  DBLP  /  Instagram

Multi-modal LLM for AutoDrive Scenarios

Generative Trajectory Planning for End-to-end Autonomous Driving

We (Flappy Pig) are the top 3 teams (175 teams in total) on this challenge. The evaluation metric is DICE value in Image Segmentation. Our DICE value is 0.902628848 (Top-1: 0.903214619). The base model is CaraNet updated by Adam optimizer.

To refine the noisy labels of noisy clients via an end-to-end optimization framework in the scenario of distributed learning.

An early attempt to deal with noisy labels considering privacy concern of Federated Learning.

Semantic Visual-inertial SLAM for Autonomous Parking

To tackle distributed long-tailed data.

To tackle both long-tailed & noisy data issue in the context of Federated Learning.

To deal with label noise in distributed health-care human activity recognition datasets, which can be seen as a time-series task.

To tackle long-tailed data issue in the context of Federated Learning.

To utilize cloud-edge-client compuatation to faciliate training larger model agglomeratively.

To tackle non-IID data issue in the context of Federated Learning with Knowledge Cache Architecture.

To deal with non-IID issue of Federated Learning with Knowledge Distillation.

To reduce communication cost in Federated Learning system via Adaptive Bayesian Dropout.

To tackle non-IID data issue in Federated Learning with lower computation cost.

My first work conducted at LiAuto AutoDrive. We achieve the state-of-the-art on the NAVSIM benchmark.

To comprehensively benchmark sixteen existing methods on five datasets and inspect why noisy labels impair federated learning.

To build robust image retrieval system against the noisy labels.

To tackle distributed noisy datasets with dual optimization strategies.

How LLMs perform for the code vulnerability detection task. One most contribution lies in the open-source implementation to support training 3 Graph-based models, 2 Bert-based sequence models and 4 open-source LLMs.

How to generate high-quality samples with VQ-GAN to tackle imbalanced distributed datasets.

To deal with Non-IID and label noise in FL via sampling scheme.

To generate diverse speakers with different timbre with your prompt via latent diffusion.

An explict yet effective method to exploit ensemble distillation to tackle Non-IID issue in Federated Learning.

A follow-up work based on FedLSR.

Incremental learning in FL with theoretical analysis.

Summary of federated distillation in edge computing environments.

Derived from Jon Barron's website.