Human Side of ML

smooth failing

Team Structure

cross functional teams collaboration, SMEs(Subject Matter Expertise)

End-to-End Data Scientist

• option 1 : have separate team to manage prod
• option 2 : Data Scientists own the entire process

  Responsible AI

• case 1 : automated grader’s biais
  • failed to set goal
  • no fine-grained eval to discover biais
  • lack of transparency
• case 2 : military data

  Framework for Responsible AI

• discover sources for model biaises
• understand limitations of data-driven approach
• understand trade-offs between desirata, e.g: compactness vs fairness
• act early
• create model cards
• establishing processes for mitigating biaises (AI Fairness 360 by IBM)
• stay updated on responsible AI