Breaking black-box AI
summary
Live demos use scikit-learn, TensorFlow and Azure Cognitive Services to show concrete failures: a single outlier turns a random dataset into a model with R² ≈ 98% and p ≈ 0, decision trees explode on diagonal decision boundaries producing uninterpretable rules, and Azure Cognitive Services mislabels unexpected objects and ignores out-of-context items. Cross-domain examples include a pneumonia classifier that learns shoulder tags, a melanoma tool that learns surgeons marks, and an Uber LIDAR pipeline that fails on jaywalking — models break because of biased training data.
AI-written summary · about
chapters (8)
- 0:00Intro
- 3:00R demo
- 8:27Decision trees
- 15:57Deep learning
- 19:15Modifying
- 20:36What does it see
- 24:34The problem
- 27:35Medicine
related talks
52mBreaking black-box AI
47mBreaking Black-Box AI
45m"Breaking black-box machine learning"
01h 00mLearn Live: Get started with AI on Azure
38mReal Life Customer Stories with AI
24mDon't want to miss out on the AI wave?
53mImplementing AI using Azure Machine Learning Services
30mTrustworthy AI - opening up the black box