MLOps | MLOps의 필요성, 개념과 MLOps 관련 상용 서비스(MakinaRocks Link)

Introduction to MLOps

• 서울대학교 컴퓨터공학부 데이터마이닝개론 특강(2022학년도, 강유 교수님)을 정리한 내용입니다.

강연자 소개

• Yongsub Lim
  • co-founder & chief data scientist of MakinaRocks
  • Yongsub Lim
  • Making industrial technology intelligent

목차

• advance of AI/ML Capabilities
• Why MLOps
• What is MLOps
• When MLEs use MLOps
• MakinaRocks Link with Demo

AI/ML의 진보

Number of Publications in NeurIPS

10년 동안 6배 이상 증가함

Accuracy of Image Classification

Image Classification에서 Human Level을 넘어섬

Investment on AI

Global investment in AI jumps to record high

Datasets, Computing Power, DeepLearning

• Imagenet
• coco - Common Objects in Context
• GPT-3

AI is Everywhere

big tech company에서 활발히 자신의 비즈니스/서비스에 적용하고 있음

• content recommendation
• language translation
• AI drawing,
• 샌프란시스코 cruise,
• AI coding(github copilot)

Why MLOps

Many technologies are being developed
As the number of successful cases increases, more companies are attempting to implement AI
(ex, chatbots, logistics optimization, EV optimization, battery management)

ML code -> 서비스/비즈니스에 적용하고 싶음.

그 외에 필요한 내용

• data collection
• data verification
• configuration
• automation
• feature engineering
• metadata management
• monitoring
• serving infrastructure
• process management
• model analysis
• resource management
• tesing & debugging

not all companies have sufficient AI, SW capabilities

AI Adoption is Challenging

(McKinseyHM19) almost 2 in 3 companies that are adopting digital manufacturing solution find theirself stuck at pilot stage

MLOps

• who does that?
• what if we need to update the ML code?
• how to ensure a new model is better than an old one?

what is MLOps?

• Machine Learning Operations
• ML
  • Experimentation
    • Data acquisition
    • Exploratory data analysis
    • Initial modeling
  • Development
    • Modeling & Testing
    • CI/CD
  • Operation
    • Model management
    • CT
• ML + DEV + OPS

Level of MLOps

• Level 0 - Manual Process
• Level 1 - Pipeline deployment, CI/CD
• level 2 - Pipeline CI/CD

Level 0 MLOPs

• operation과 분리되어 진행되는 model development

-> A frequent Problem
Deployment side와 ML side가 충돌함

Level 1 MLOps

• feature store
• pipeline deployment
• continuous training
• CD : Model Serving

Level 2 MLOps

• More CI/CD for Pipelines

More on MLOps

Google "MLOps:Continuous delivery and automation pipelines in machine learning"
https://cloud.google.com/architecture/mlops-continuous-delivery-and-automation-pipelines-in-machine-learning?hl=ko
모두의 mlops

Many Tools on MLOps

Kubeflow
Azure Machine Learning ...

Our experience is

• Still, there is huddles for data scientist
  DS/MLEs are good at -> Data analysis, Model analysis, Orchestrated experiment
  model 결과를 deployable, structured program으로 바꾸어야 한다.
• Deployed pipeline needs debugging

MakinaRocks Link with Demo

Convert and share ML models developed by data scientists, in a pipeline format in the Ops environment

• ML pipeline creator

  • convert ML codes written by DS in Jupyter Notebook
  • to Pipelines operable in MLOps environment

• Why MRX Link?

  • Remove technical hurdles for usage in MLOps
  • Better collaboration between DS
  • Improved efficiency