The Path to Machine Learning

Steps to Turn Process into Machine Learning Model

Individual Contributor

• Prototype
• Don’t skip—leads to incorrect assumptions later
• Don’t linger—knowledge can become concentrated in 1 person
• Slow to scale

Delegation

• Increase number of employees—100s to millions
  • But—tool many stakeholders can result in organizational lock-in—“automation not possible”
• Formalize process
• Org buy in
• Identify humans in ML loop

Digitization

• Computers perform mundane tasks
• Automation
• Lay software and infrastructure foundations—ML extra layer on top

Big Data and Analytics

• Measure everything about ops and users
• Use data to reassess definition of success
• Hand-tuned algorithms—e.g. Google until 2015

Machine Learning

• Complete feedback loop—automate all steps

Moving from Experimentation to Production

• Frame the problem
  • Identify the use case
  • What?
  • Minimum requirements?
• Prepare training data
  • Exploratory data analysis—EDA
    • Improve data quality
  • Feature engineering—combine features to create new features
• Experimentation phase
  • Experiment with different models
• Evaluate
  • Precision, recall, F1 score, cross entropy
• Package, deploy, monitor