Jiatao Gu
Assistant Professor at University of Pennsylvania
New York, New York, United States
Summary
🤩
Rockstar🎓
Top SchoolJiatao Gu is an Assistant Professor at UPenn and a Staff Research Scientist at Apple AI/ML with 12 years of experience building generative, multi-modal machine learning systems that bridge language, vision, video and 3D. He previously led research at Meta AI (FAIR) and contributed core improvements to the widely used fairseq toolkit—advancing non‑autoregressive transformers, CUDA-optimized Levenshtein decoding, and length beam search—highlighting both research depth and production-aware engineering. Trained at Tsinghua and HKU (Ph.D.), his work focuses on world modeling, iterative reasoning, and decision-making for AI agents operating in complex physical environments. Known for combining foundational ML research with practical optimizations, he aims to make models that are efficient, flexible, scalable and knowledge-rich. Based in New York, he blends academic rigor with industry impact, often taking on low-level implementation challenges that accelerate model deployment.
12 years of coding experience
4 years of employment as a software developer
University of TokyoThe University of Hong Kong (HKU)Bachelor's degree Electronic Engineering, Bachelor's degree Electronic Engineering at Tsinghua University
English, Japanese, Chinese
Github Skills (13)
transformer-models10
pytorch10
machine-learning10
artificial-intelligence10
python10
fairseq10
natural-language-processing9
nlp9
cuda9
algorithms8
data-structures8
algorithm8
data-structure8
Programming languages (6)
Github contributions (5)
facebookresearch/fairseq
Oct 2019 - Jan 2020
Facebook AI Research Sequence-to-Sequence Toolkit written in Python.
Role in this project:
ML Engineer
ML EngineerContributions:21 commits, 58 comments, 1 issue in 3 months
Contributions summary:Jiatao's primary contributions involve enhancing and refining the fairseq library, with a focus on Non-Autoregressive Transformers (NAT) and related models. This is evident in the implementation of new functionalities, like the "new_arange" function and correcting bugs in the returning attention values. Further work includes refactoring NAT implementations, incorporating CUDA optimizations for Levenshtein distance calculations, and integrating advanced decoding strategies such as length beam search. The user is working on core improvements to support sequence-to-sequence modeling with transformers.
aipythonsequence-to-sequencepytorchartificial-intelligence
MultiPath/SEG-NMT
Feb 2017 - Oct 2017
Contributions:231 commits in 8 months
Find and Hire Top DevelopersWe’ve analyzed the programming source code of over 60 million software developers on GitHub and scored them by 50,000 skills. Sign-up on Prog,AI to search for software developers.