Bias, Explainability, and Accountability: The Data Scientist's Burden
summary
Examples of real failures (Tesla crash 2018, Apple card gender bias, COMPAS showing 23% vs 45% recidivism) motivate concrete fixes: prefer decision trees or logistic regression when they meet metrics, apply SHAP and LIME to justify black‑box models, generate feasible counterfactuals, run shadow deployments, and audit imbalances (adult income: males ~2× females; male 1/3 >$50k vs female 1/5). Adopt FACT: Fairness, Accuracy, Confidentiality, Transparency and use bias‑aware trees.
AI-written summary · about
chapters (21)
- 0:00Introduction
- 0:43Title and ethical AI
- 1:28Impact on society
- 2:53Irresponsible data science
- 3:57Bias in AI systems
- 5:26Statistical vs societal bias
- 6:31Interpretability vs explainability
- 8:19Complex models and explainability
- 9:45Explainable AI techniques
- 11:51Counterfactual explanations
- 13:17Racial bias in risk assessment
- 14:21Gender imbalance in income data
- 15:02FACT standards
- 15:43Ethical burden and accountability
- 17:14Closing thoughts
- 17:36Q&A
- 18:59Model complexity vs explainability
- 22:09Data imbalance and bias
- 23:58Ending the bias cycle
- 27:10Overcoming math barriers
- 29:18Closing remarks
related talks
21mFairness and bias in AI
27mCrouching Harm, Hidden Bias—Addressing issues of fairness....
38mSocial Implications of Bias in Machine Learning
25mBias in Data
54mEthical Machine Learning
31mFlagging *and* Fixing Bias in ML
38mSocial Implications of Bias in Machine Learning
40mQuestions and Practices to Make Algorithmic Decision Making More Fair