"Breaking black-box machine learning"
summary
Shows concrete failures and tools: uses R linear regression, scikit-learn decision trees, F# and Microsoft Cognitive Services vision API on real datasets (bike sharing, Titanic, donorchoose, StackOverflow survey). Demonstrates one outlier turns R² 0% → 98% and a single bad temperature entry drops bike-sharing R² from ~80% to ~73%, exposes that linear models fail on standardized-error outliers, that decision trees fragment a simple sum rule into many split tests, and that the vision API mislabels unusual combos (monkey holding guitar becomes dog).
AI-written summary · about
chapters (5)
- 0:00Introduction to black box ML
- 5:57Linear regression risks
- 17:25Decision tree limitations
- 27:40Deep learning and vision APIs
- 37:55Data integrity and conclusion
related talks
52mBreaking black-box AI
47mBreaking Black-Box AI
35mBreaking black-box AI
53mUnleashing the Black Box: Testing Machine Learning Systems
52mMachine Learning Crash Course
34mInterpreting Machine Learning Models: Why and How!
01h 02mA developers guide to Machine Learning
34mData Science beasts (failures) and where to find them