HE Xin (贺鑫), a.k.a. MARSGGBO
AI Research Fellow, National University of Singapore
![ggbo_on_mars.jpg](/assets/img/ggbo_on_mars.jpg?b17e958b801cfcf8fa1cda303450183f)
hexin@nus.edu.sg
| 知乎 | Wechat | 博客园 | 腾讯云+社区 | Source code |
Driven by a mission to democratize deep learning, my research is dedicated to advancing the accessibility and efficiency of large-scale deep learning models, particularly Large Language Models (LLMs). My goal is to bridge the theoretical aspects of machine learning with practical system designs to create scalable, robust, and trustworthy AI systems that are widely accessible. My interested research directions include:
- 1.Model-Centric AI:
- Architecture Dearch: Neural Architecture Search (e.g., multi-objective NAS, Training-free NAS, resource-aware NAS), Sparse Model (e.g., Mixture-of-Experts)
- Hyper-parameter optimization (HPO): Grid/Random Search, Evolutionary Algorithm, Differentiable Optimization
- Model Compression: Pruning, Quantization, Knowledge Distillation
- 2.Data-Centric AI:
- Automatic Data Augmentation (ADA), Data Generation, Dataset compression,
- RAG, LLM alignment
- 3.HPC AI:
- Memory efficiency: Offloading, KV-cache
- LLM training acceleration: Distributed Parallellism (data parallel, tensor parallel, pipeline parallel)
- LLM inference optimization: Batch Schedule, Dynamic Inference Paths
Education & Work
- 🇸🇬 National university of Singapore (NUS), Singapore (2023-now)
- Research Fellow, School of Computing
- Principle Investigator: Prof. You Yang
- 🇭🇰 Hong Kong Baptist University (HKBU), Hong Kong, China (2018 - 2023)
- Ph.D, Department of Computer Science
- Supervisor: Prof. Chu Xiaowen
- 🇨🇳 Huazhong University of Science & Technology (HUST), Wuhan, China (2014 - 2018)
- Bachelor, School of Electronic Information and Communications, Honor of Outstanding Student
Interested in collaboration? Contact me.
Highlights
- 1.AutoML Survey (1400+ citation): AutoML: A Survey of State-of-ther-art
- 2.AutoML Framework: Hyperbox
- 3.AutoML Applications:
- AAAI2021 COVID3DNet: The first NAS application for COVID-19 3D CT scans
- MICCAI2022 EMARS: Evolutionary algorithm-based NAS for COVID-19 3D CT scans
- ECCV2022 EAGAN: Two-stage NAS for GANs
- AAAI2023 NAS-LID: NAS with Local Intrinsic Dimension
news
Feb 23, 2024 | I am honored to serve as a reviewer for ECCV’24. |
---|---|
Sep 03, 2023 | Our paper “MedPipe: End-to-End Joint Search of Data Augmentation Policy and Neural Architecture for 3D Medical Image Classification” is accepted to IEEE MedAI2023! |
Aug 09, 2023 | I am honored to serve as a reviewer for AAAI’24. |
Nov 19, 2022 | Our paper “NAS-LID: Efficient Neural Architecture Search with Local Intrinsic Dimension” is accepted to AAAI2023! |
Nov 02, 2022 | I am honored to serve as a reviewer for CVPR’23. |
latest posts
Jun 26, 2024 | Deepspeed ZeRO系列算法原理+通信开销详解 - marsggbo |
---|---|
Jun 09, 2024 | NSCC集群使用笔记 - marsggbo |
May 06, 2024 | Huggingface Transformers实现张量并行的小坑 set/get_output_embeddings - marsggbo |