HyunJae Lee
Machine Learning Engineer
I am a full-stack machine learning engineer with extensive experience in both research and engineering. My research has consistently focused on solving real-world problems, which has resulted in multiple publications at top-tier conferences. I have established and expanded an MLOps team from scratch to six members, and I led the design and implementation of both a cost-effective, cloud-based ML platform and an automated AI model development framework. I was responsible for the full system design and the initial development, handling backend, frontend, and DevOps all at once. I am passionate about solving real-world problems by bridging the gap between ML research and engineering.
Work Experience
Lunit Inc.
Seoul, South Korea
Research Engineer / AutoML Team Leaader, July 2021 – Present
- Established and scaled the team to six members focused on automation and efficiency in ML model development.
- Architected and implemented a large-scale, cost-efficient ML platform, leveraging over 1000 GPUs for simultaneous training at a cost reduction of more than 50%. This platform supports nearly 100,000 experiments annually.
- Designed and developed an automated AI model development framework that streamlines the repetitive aspects of AI model creation, enhancing efficiency and consistency.
- Advanced research in Bayesian optimization and pruning algorithms for hyperparameter optimization, resulting in publications and direct integration into our operational framework.
Research Scientist / 3D Mammography Team Leader, March 2018 – July 2021
- Conducted research that addresses the domain gap in both internal and external datasets, significantly enhancing model performance and robustness, leading to publications in top-tier conferences and practical application.
- Developed a 3D Mammography cancer detection AI model that received FDA 510(k) clearance.
- Built a complete pipeline from scratch, including data preprocessing, model development and model deployment.
ConvIoT
Seoul, South Korea
Lead Developer / CEO, June 2016 – July 2017
- Developed and launched a user-friendly IoT platform enabling simple workflow configuration via a web interface.
- Integrated a wide range of internet services and IoT devices from diverse vendors, enabling seamless and automated interactions across systems, and secured early investment from Naver.
Selected Publications (700+ citations)
Bayesian Optimization Meets Self-Distillation
HyunJae Lee*, Heon Song*, Hyeonsoo Lee*, Gi-hyeon Lee, Suyeong Park, Donggeun Yoo
I developed the BOSS framework, which enhances model performance by combining Bayesian Optimization (BO) and Self-Distillation (SD). Unlike traditional BO, which only uses partial knowledge from previous trials, BOSS integrates insights from all training trials. This innovative approach not only suggests optimal hyperparameter configurations but also utilizes pre-trained models for self-distillation, leading to superior performance across various tasks such as image classification, noisy label learning, semi-supervised learning, and medical image analysis.
Reducing Domain Gap by Reducing Style Bias
HyunJae Lee*, Hyeonseob Nam*, Jongchan Park, Wonjun Yoon, Donggeun Yoo
I developed Style-Agnostic Networks (SagNets) to address the domain shift problem in Convolutional Neural Networks (CNNs). SagNets reduce the intrinsic style bias of CNNs by disentangling style encodings from class categories, allowing the networks to focus more on content rather than style. This approach enhances the robustness of models across different domains. Extensive experiments demonstrate that SagNets significantly improve performance in various cross-domain tasks, including domain generalization, unsupervised domain adaptation, and semi-supervised domain adaptation.
SRM : A Style-based Recalibration Module for Convolutional Neural Networks
HyunJae Lee, Hyo-Eun Kim, Hyeonseob Nam
I introduced the Style-based Recalibration Module (SRM) to enhance the performance of Convolutional Neural Networks (CNNs) in various vision tasks by leveraging style information. SRM adaptively recalibrates feature maps based on their styles, improving the representational capacity of CNNs. This module integrates seamlessly into existing architectures with minimal overhead. Extensive experiments demonstrate that SRM outperforms recent methods, including Squeeze-and-Excitation (SE), in tasks such as general image recognition and style-related tasks. An in-depth comparison of SRM and SE highlights their distinct representational properties.
Featured Projects
AutoSCOPE Framework
AutoSCOPE is an automated AI product development framework designed to streamline and automate the creation of deep learning models within Lunit's oncology department. It simplifies the entire AI model lifecycle, from data management and model training to inference on whole slide images, through a user-friendly web interface. By automating complex, repetitive tasks and integrating advanced AutoML algorithms, AutoSCOPE enhances efficiency, reduces the need for extensive cross-team communication, and empowers domain experts without AI expertise to directly contribute to model development. This results in faster, more scalable, and clinically relevant AI solutions in oncology research.
Intelligent Cloud Framework
INtelligent CLoud (INCL) is a deep learning training platform developed by Lunit to streamline and optimize the process of training AI models on the cloud. INCL automates manual processes involved in cloud-based model training, including experiment setup, tracking, and management, while providing advanced features like multi-node distributed learning and automated hyperparameter optimization. This platform significantly enhances scalability, reduces the likelihood of errors, and allows researchers to focus on model development rather than infrastructure management. By leveraging INCL, Lunit has been able to efficiently scale its computational resources, improve model performance, and accelerate the overall research process.
ConvIoT
ConvIoT is an innovative Internet of Things (IoT) platform designed to seamlessly connect and automate various devices and services, enhancing user convenience and simplifying daily tasks. By providing a user-friendly web interface, ConvIoT allows users to easily create and manage applets that integrate different IoT devices and web services, enabling automation with just a few clicks. The platform supports a wide range of devices, offers robust security features, and generates necessary code for device integration, making it accessible to both consumers and manufacturers. ConvIoT's versatility and ease of use maximize the utility of connected devices, offering significant potential for both user convenience and vendor marketing strategies.
Selected Awards
1st place in Visual Domain Adaptation Challenge
ICCV 2019
Won first place in the famous VisDA-2019 Challenge as the main contributor, demonstrating excellence in Semi-Supervised Domain Adaptation and employing techniques to address real-world challenges in the medical domain.
2nd place in Embedded Deep Learning Design Challenge
ESWeek 2017
Achieved second place in the EDLDC 2017, demonstrating an energy-efficient deep learning design for object detection on embedded systems, utilizing advanced techniques such as reduced precision inference, tensor decomposition, etc.
Best Start-up Award
Naver Demo Day 2017
Won the Best Start-up Award at Naver Demo Day 2017 as Lead Developer/CEO of ConvIoT, for developing and launching an innovative IoT platform that simplifies workflow configuration and integrates various internet services and devices.
Technical Skills
ML & MLOps
Pytorch, Scikit-learn, Optuna, Vertex AI, MLflow, Kubeflow
Backend Development
Django, DRF, FastAPI, Gin, Celery, Firebase, MySQL, PostgreSQL, Redis, Kafka
DevOps & Infra.
Kubernetes, Docker, Terraform, ArgoCD, GitHub Actions, Prometheus, Grafana, Loki, GCP, AWS
Frontend Development
React, Next.js, MobX, Vercel, Material UI, Tailwind CSS, Shadcn
Languages
English (fluent), Korean (native)