Breaking black-box AI
summary
Demos run with R, scikit‑learn and Microsoft Cognitive Services expose concrete failures: R linear regression: one outlier raises R² from ~2.5% to 98% and makes p≈0, decision trees carve fine grids on a simple diagonal dataset and become unreadable, APIs label surreal composites as dogs or miss an elephant in an office, and medical nets learn shortcuts — pneumonia model looks at a shoulder tag and skin lesion classifier flips to malignant 100% when surgical marks or scale change. The problem is biased, messy training data, not mysterious models.
AI-written summary · about
chapters (5)
- 0:00Start
- 11:41DEMO LINEAR REGRESSION
- 12:38HUMAN COMPUTER INTERACTION
- 13:38DECISION TREE: WHO SURVIVES ON TITANIC
- 21:15DEMO DEEP LEARNING
related talks
35mBreaking black-box AI
47mBreaking Black-Box AI
45m"Breaking black-box machine learning"
30mTrustworthy AI - opening up the black box
31mThe dark side of deep learning
34mInterpretable AI or How I Learned to Stop Worrying and Trust AI
49mAn AI with an Agenda
56mArtificial Intelligence? - more like Artificial Stupidity!