Chankyu Lee

Chankyu Lee

Senior Research Scientist

NVIDIA

Biography

Chankyu Lee is a Senior Research Scientist at NVIDIA's Applied Deep Learning Research (ADLR) team, where he focuses on advancing large language models, including post-training research for embedding, reasoning, and agentic coding models. He obtained Ph.D. degree from Electrical and Computer Engineering, Purdue University (Advisor: Prof. Kaushik Roy).

Interests

  • Large Language Models
  • Agentic AI
  • Algorithm-Hardware Co-design

Education

  • PhD in Electrical and Computer Engineering, Fall 2015 - Spring 2021

    Purdue University, West Lafayette, IN, USA

  • BS in Electrical and Electronics Engineering, Spring 2009 - Spring 2015

    Sungkyunkwan University (SKKU), Suwon, South Korea

  • Exchange Student Program, Electronic and Computer Engineering, Fall 2013

    Hong Kong University of Science and Technology (HKUST), Hong Kong

Experience

 
 
 
 
 

Senior Research Scientist

NVIDIA Corporation

Mar 2022 – Present Santa Clara, California, USA

Applied Deep Learning Research

  • Post-training LLM for reasoning capability (agentic coding, code generation, math, etc). Part of Ace-reason and nemotron-cascade effort, best-in-class 8B, 14B and 30B-A3B LLM. Extended to flagship nemotron model family (nano/super/ultra).
  • Information Retrieval: embedding model and Retrieval Augmented Generation (RAG). NV-Embed-v1 and v2: No. 1 ranking embedding models on the MTEB leaderboard and 2M huggingface model downloads. MM-Embed: multimodal extension of NV-Embed.

TAO-toolkit and autoML

  • Accelerating the model training process by abstracting away the AI/deep learning framework complexity [link].
 
 
 
 
 

Largescale Machine Learning Engineer

Intel Corporation

Apr 2021 – Mar 2022 Austin, Texas, USA
AI workload performance optimization (i.e., improving the runtime, efficiency, scalability) for the largescale machine learning training on Intel AI accelerators (Gaudi-Habana Labs) [link].
 
 
 
 
 

Graduate Internship

Bell Labs, Nokia

Jun 2018 – Aug 2018 Murray Hill, New Jersey, USA
Developed the functional modeling simulator for mapping and scheduling CNN (AlexNet, VGG, ResNet) algorithms on a MIMD (Multi-Instruction Multi-Data) processor toward accelerated and energy-efficient AI computing.
 
 
 
 
 

Graduate Research Assistant

Purdue University

Aug 2016 – Dec 2020 West Lafayette, Indiana, USA

Exploratory research on neuromorphic computing for energy-efficient and robust deep learning, overcoming limitations of current artificial intelligence through algorithm-hardware co-design.

  • Topic 1: Developed novel unsupervised/supervised/semi-supervised learnings for deep convolutional Spiking Neural Networks (SNNs) to efficiently harness machine learning algorithms.
  • Topic 2: Developed energy-efficient motion estimation algorithms for event-based camera in challenging scenes such as high-speed and wide-dynamic range.
  • Research Outputs: 13 publications including 6 first-authored papers (1 IEEE TCDS, 2 Frontiers in Neuroscience, 1 ECCV, 1 Neurocomputing, 1 IEEE ICRA).
 
 
 
 
 

Undergraduate Research Assistant

Graduate School of Convergence Science and Technology, Seoul National University

Jul 2014 – Aug 2014 Suwon, South Korea
Research on mid-field wireless powering for simulating neural signals in the brain-machine interface. Advised by Prof. Yoonkyu Song

Selected Publications

Nemotron-Cascade: Scaling Cascaded Reinforcement Learning for General-Purpose Reasoning Models

Scaling cascaded reinforcement learning for general-purpose reasoning models, achieving best-in-class 8B, 14B and 30B-A3B LLMs.
Boxin Wang*, Chankyu Lee*, Nayeon Lee*, Sheng-Chieh Lin*, Wenliang Dai*, Yang Chen*, Yangyi Chen*, Zhuolin Yang*, Zihan Liu*, Mohammad Shoeybi, Bryan Catanzaro, Wei Ping. Technical Report (NVIDIA), 2025
PDF
Nemotron-Cascade: Scaling Cascaded Reinforcement Learning for General-Purpose Reasoning Models

NV-Embed: Improved Techniques for Training LLMs as Generalist Embedding Models

Improved techniques for training LLMs as generalist embedding models, achieving No. 1 ranking on the MTEB leaderboard.
Chankyu Lee, Rajarshi Roy, Mengyao Xu, Jonathan Raiman, Mohammad Shoeybi, Bryan Catanzaro, Wei Ping. International Conference on Learning Representations (ICLR spotlight), 2025
PDF
NV-Embed: Improved Techniques for Training LLMs as Generalist Embedding Models

Fusion-FlowNet: Energy-Efficient Optical Flow Estimation using Sensor Fusion and Deep Fused Spiking-Analog Network Architectures

Standard frame-based cameras that sample light intensity frames are heavily impacted by motion blur for high-speed motion and fail to …
Chankyu Lee, Adarsh Kosta, Kaushik Roy. IEEE International Conference on Robotics and Automation (ICRA), 2022
PDF
Fusion-FlowNet: Energy-Efficient Optical Flow Estimation using Sensor Fusion and Deep Fused Spiking-Analog Network Architectures

Towards Understanding the Effect of Leak in Spiking Neural Networks

Spiking Neural Networks (SNNs) are being explored to emulate the astounding capabilities of human brain that can learn and compute …
Sayeed Shafayet Chowdhury*, Chankyu Lee*, Kaushik Roy (*Equal Contribution). Neurocomputing, 2021
PDF
Towards Understanding the Effect of Leak in Spiking Neural Networks

Spike-FlowNet: Event-based Optical Flow Estimation with Energy-Efficient Hybrid Neural Networks

Event-based cameras display great potential for a variety of conditions such as high-speed motion detection and enabling navigation in …
Chankyu Lee, Adarsh Kosta, Alex Zihao Zhu, Kenneth Chaney, Kostas Daniilidis, Kaushik Roy. In Proceedings of the European Conference on Computer Vision (ECCV) 2020
PDF Code Video
Spike-FlowNet: Event-based Optical Flow Estimation with Energy-Efficient Hybrid Neural Networks

Enabling Spike-Based Backpropagation for Training Deep Neural Network Architectures

Spiking Neural Networks (SNNs) have recently emerged as a prominent neural computing paradigm. However, the typical shallow SNN …
Chankyu Lee*, Syed Shakib Sarwar*, Priyadarshini Panda, Gopalakrishnan Srinivasan, Kaushik Roy (*Equal Contribution). Frontiers in Neuroscience, Neuromorphic Engineering, 2020
PDF Code
Enabling Spike-Based Backpropagation for Training Deep Neural Network Architectures

Training Deep Spiking Convolutional Neural Networks with STDP-based Unsupervised Pre-training Followed by Supervised Fine-tuning

Spiking Neural Networks (SNNs) are fast becoming a promising candidate for brain-inspired neuromorphic computing because of their …
Chankyu Lee, Gopalakrishnan Srinivasan, Priyadarshini Panda, Kaushik Roy. Frontiers in Neuroscience, Neuromorphic Engineering, 2018
PDF
Training Deep Spiking Convolutional Neural Networks with STDP-based Unsupervised Pre-training Followed by Supervised Fine-tuning

Contact