Full Stack Deep Learning Lecture 5 ~ 7

Lecture 5

Setting up ML Projects

85% of AI projects fail

 

Prioritizing projects

ML project의 feasibility

Data availability

• stable?
• hard to acquire?
• expensive labeling?
• how much data is needed?
• security requirements?

Accuracy requirement

• how costly are wrong pred?
• how frequently does the system need to be right to be useful?
• ethical implications

Problem difficulty

• is the problem well-defined?
• good published work on similar problems?
• compute requirements
• Can a Human do it?

Apple의 ML product 디자인 가이드

• What role does ML play in your app?
  • Critical or complementary?
  • Private or public?
  • Proactive or retroactive?
  • Visible or invisible?
  • Dynamic or static?
• How can you learn from your users?
  • Explicit feedback
  • Implicit feedback
  • Calibration during setup
  • Corrections

How to combine metrics

• Simple weighted avg
• Threshold n-1 metrics, evaluate the nth

Thresholding metrics

• which metrics?
  • 도메인에 따라 결정,
  • 모델의 선택에 가장 덜 민감한 metrics
  • 얻고자 하는 결과에 가장 근접한 metrics
• threshold values
  • 도메인에 따라 결정
  • baseline의 성능에 따라 조정
  • 현재 metric이 얼마나 중요한 지를 고려

mAP (mean Average Precision)

AP: Recall - Precision 그래프의 하단 면적

 

 

Lecture 6

• Use python
  • scientific and data computing libraries

Editors

• Vim
• Emacs
• Jupyter
  • Great as first draft
  • hard to scale up
  • Hart to version
  • not good for distributed task, long task
• VS Code
  • peek docu
  • bulit-in git tolls
  • remote
  • • terminal
  • notebook port forwardin
• PyCharm

Streamlit

• python code → visualization

 

Framework

Why?

• auto-differentation and CUDA are a lot of work

Lecture 7

남은 강의들은 노션에 정리 예정.

 

 

Reference 

[1] Full Stack Deep Learning 5~7
https://www.youtube.com/playlist?list=PL1T8fO7ArWlcWg04OgNiJy91PywMKT2lv