Breaking Black-Box AI
summary
Uses R and scikit-learn to show concrete failures: adding one outlier at (100,100) turns a 90% p-value random fit into a model with 98% R‑squared, a single mistaken Fahrenheit entry in a 700‑row bike dataset drops performance by 7% R‑squared, and a simple two‑feature rule explodes into a huge, unreadable decision tree. Calls Azure Cognitive Services to show image tags flip (guitar → dog, hidden apple ignored). Cites medical papers where metal tags and surgical marks make classifiers predict pneumonia or malignancy. black‑box models learn dataset quirks.
AI-written summary · about
chapters (5)
- 0:00Intro
- 1:40MACHINE LEARNING APIS
- 14:03DEMO DECISION TREE
- 22:35DEMO DEEP LEARNING
- 33:09MACHINE VISION MEDICINE
related talks
35mBreaking black-box AI
52mBreaking black-box AI
45m"Breaking black-box machine learning"
37mThe AI 'Black Box'
30mTrustworthy AI - opening up the black box
29mOpeningTheBlackBox
39mKeynote 3: How did you know? Explaining Black Box Model Predictions in Python
04mBlack Box: How AI will Amplify the Best and Worst of Humanity