Toward Human-Centered Machine Learning
summary
Describes a reproducible workflow that combines GAMs and scalable Bayesian rule lists with post-hoc tools like Shapley values and LIME to produce interpretable, fair, and monitorable models; shows a credit model that overweights the most recent payment producing large residuals and logical failures; warns about attacks (data poisoning, adversarial examples, membership inference), cites Shapley from 1953 applied to ML in 2017, recommends disparate-impact checks (20% rule), rigorous model debugging, deployment monitoring, and appealable stored models.
AI-written summary · about
chapters (17)
- 0:00Introduction
- 2:20The Blueprint
- 2:52Understand Your Data
- 7:15Record Your Data
- 7:44Feature Engineering
- 9:00Privacy
- 9:55AccuracyInterpretability
- 10:48Residual Analysis
- 12:00Explanations
- 15:26Shapley Values
- 18:47Trust and Understanding
- 23:39Model Debugging
- 25:59Machine Learning Attacks
- 29:26Dis disparate impact
- 33:15Document and review models
- 35:29Appeal decisions
- 39:29Improving accuracy
related talks
30mMachine Learning Techniques for Class Imbalances & Adversaries
27mInterpretability and explainability in machine learning
28mManaging AI Risk: Beyond Explainability
50mEthical Machine Learning
47mEthical Machine Learning
34mInterpreting Machine Learning Models: Why and How!
56mDehumanizing Agents: Why Explainability is Crucial in the LLM Era
39mBeyond Bias: Crafting Fairness in Pythonic Machine Learning Systems