AIM215-R1 – Introducing SageMaker Autopilot (re:Invent 2019) – Key Takeaways

The Key

  • Scaling ML != training single models, but includes searching in combinations of different preprocessors, models, environments, datasets, hyperparameter tuning
  • Autopilot features automated optimizations of
    • Hyperparameters
    • Instance types
    • Preprocessors
    • Algorithms

The Takeaways

  • Autopilot is not Automated ML
    • Blackbox, non-auditable
    • Overly simplified for non-experts
    • Difficult to tune and make tradeoffs
  • Autopilot changes this
    • Generates notebooks for inspection, learning, tuning and auditing
    • Select automatically or deploy all models
    • Still provide simple access
  • Screenshots
    • UI integrated into SageMaker Studio, or use Python to access programmatically
    • Notebooks generated by Autopilot
      • Data Exploration Notebook
      • Candidate Definition Notebook
        • Available Knobs
  • DevFactory
    • NOTE: This guy talks okay, but the slides were text-based with very little information
    • Slogan: Take data to cloud and bring new values
    • Clients
      • ZephyrTel: churn prediction
        • from traditional RPM to 40 vectors
      • Ignite Technologies: intrusion and anomaly detection
        • from rule-based to AI-driven
        • detect new unseen attacks
      • SmartLeads
        • lead scoring engine
        • sub 100ms latency

The Non-essentials

  • Data explosion, ML basics, challenges
    • Demo
      • a very brief demo of Autopilot in SageMaker Studio, not all that informational