Blog Posts

AutoSCOPE: The Assembly Line of AI Model Development at Lunit

AutoSCOPE is an advanced AI development framework designed to revolutionize the process of building AI models, particularly within the field of oncology. By automating and simplifying complex tasks, AutoSCOPE streamlines the entire AI model development lifecycle—from dataset management to whole slide image inference—making it as efficient and systematic as an assembly line. This tool not only accelerates the development cycle but also improves team coordination, allowing domain experts without deep AI expertise to directly train and refine models. AutoSCOPE's user-friendly interface and advanced AutoML algorithms reduce the need for manual intervention and inter-team communication, significantly speeding up the creation of high-quality, clinically relevant AI models.

Bayesian Optimization Meets Self-Distillation

The BOSS framework combines Bayesian Optimization (BO) with Self-Distillation (SD) to significantly enhance deep learning model performance. BOSS addresses the challenges of traditional hyperparameter tuning by strategically utilizing knowledge from previously trained models, inspired by recent advancements in SD. This approach not only optimizes hyperparameters but also builds on past insights, leading to consistent performance improvements across various tasks, including image classification and medical diagnostics. By automating this process within the INtelligent CLoud (INCL) platform, BOSS dramatically improves the efficiency and reliability of AI models, ensuring they are more effective in real-world clinical applications.

Intelligent Cloud — Part 1: Introduction to Lunit’s Cloud Deep Learning Platform for Efficient Model Training

To address the challenges of scaling deep learning model training, including issues with on-premise servers and the complexity of cloud-based training, the INtelligent CLoud (INCL) platform was developed. INCL automates and streamlines the entire deep learning training process, from setting up cloud environments to running multi-node distributed learning and optimizing hyperparameters. This platform has significantly improved research efficiency by eliminating manual tasks, enhancing experiment tracking and management, and enabling large-scale model training with minimal effort. By migrating to the cloud and leveraging INCL, the research team has been able to scale up infrastructure rapidly, achieve state-of-the-art performance, and focus more on innovation rather than infrastructure management.

Intelligent Cloud — Part 2: A Deep Dive into the Architecture and Technical Details of INCL

INCL is a cloud-based deep learning platform designed to enhance scalability, performance, and user experience in AI research. Its architecture consists of three main components: the INCL client, backend, and job instances. The client provides both command-line and web interfaces for interacting with the platform, while the backend, running on a Kubernetes cluster, manages job scheduling, execution, and logging. Each job runs on a virtual machine, or job instance, which handles the execution of experiments and logs results in real-time. INCL optimizes storage usage by leveraging a combination of object, block, and file storage, balancing performance and cost.

Intelligent Cloud — Part 3: Optimizing GPU Costs by Leveraging Spot Instances

This post focuses on how the INCL platform optimizes GPU costs by effectively leveraging spot instances for deep learning model training. Spot instances, which are significantly cheaper than on-demand instances, present a cost-effective solution but come with the challenge of potential preemptions by cloud providers. INCL addresses this by automatically resuming training after preemptions, selecting between spot or on-demand instances based on empirical data to avoid endless preemption loops. Additionally, INCL efficiently manages multi-node distributed learning, ensuring minimal disruption when either master or child nodes are preempted.

What Makes a Good Research Engineer?

This post explores the qualities that distinguish a good research engineer, emphasizing the transition from theoretical research to practical application in AI. Key traits include a robust foundation in both theoretical knowledge and practical skills, the ability to identify and define impactful problems, and the development of innovative, practical solutions. Effective research engineers also excel in clarifying requirements, architecting scalable systems, and implementing solutions with clean, maintainable code. Continuous learning and adaptation to new technologies and industry shifts are crucial, along with developing a holistic view to identify and solve significant problems within an organization.