ML and technical debt

Machine Learning: The High Interest Credit Card of Technical Debt

• link to paper: https://research.google/pubs/pub43146/

• Complex Models Erode Boundaries

  • Entanglement
    • Generally not possible to make isolated changes - Changing Anything Changes Everything
    • Applies to features, signals, parameter settings, etc
    • Somewhat innate to ML
  • Hidden feedback loop
    • ML system’s predictions end up influencing its own training data
    • May happen in surprising ways, such as two systems are dependencies of each other
    • Can result in gradual changes not immediately visible, hard to detect and debug
  • Undeclared consumers
    • Changes to model impact undeclared downstream app
    • Also potential to create hidden feedback loops

• Dependency debt

  • Data dependency cost more than code dependencies
    • Data dependency is harder to track than code dependencies
    • Unstable dependency
    • Legacy features
    • Bundle features
    • episilon Features – small improvement in accuracy with huge complexity overhead
    • Correction cascade – tendency to use another model and learn a calibration layer

• System level spaghetti

  • Glue code - most of the code is not the model itself.
  • Pipeline jungles
  • Dead Experimental Codepaths
  • Configuration Debt