Interpretable AI or How I Learned to Stop Worrying and Trust AI
summary
Demonstrates concrete techniques: uses partial dependence plots to expose feature interactions and sampling gaps in a student-grades model, applies SHAP values to explain individual breast‑cancer predictions (reports near 100% average precision on the test split for a tuned ensemble and shows which features push probability), and runs adversarial debiasing on word2vec to reduce gendered analogies so professional terms appear instead of stereotyped words. Code and notebooks live on the linked GitHub.
AI-written summary · about
chapters (8)
- 0:00Interpretable AI introduction
- 2:35Human brain to model training
- 7:05Why feature importance fails
- 9:55Feature interactions with PDPs
- 12:51Post-hoc SHAP explanations
- 20:15Bias mitigation in ML
- 27:02Summary
- 33:50Q&A
related talks
46mCan you trust your AI?
30mTrustworthy AI - opening up the black box
39mKeynote 3: How did you know? Explaining Black Box Model Predictions in Python
47mML: Explainable Artificial Intelligence explained
39mThe Search for Transparency and Accountability in the Age of...
40mResponsible AI Practices: Fairness in ML
28mTrusted AI – Building Reproducible, Unbiased and...
01h 00mData, visualization, and designing AI