Keynote: Ethical Machine Learning: Creating Fair Models in an Unjust World
summary
Claims that real systems reproduce societal bias with concrete examples and fixes: ProPublica COMPAS finding shows risk scores correlate with race; Google Translate gender flips reveal word‑vector bias; Flickr30k captions encode gender and racial stereotypes; a linear SVM demo shows discrimination improves with more data; proposed tools include the fairtest Python library, WEAT for embeddings, ensembles that segregate sensitive features (Meetup), and enforcing equalized odds / equal opportunity plus maximum‑activation analysis to inspect and remove unwanted learned attributes.
AI-written summary · about
chapters (9)
- 0:00Ethical ML introduction
- 1:48Biased algorithms' real-world impact
- 7:17Injustice in data and models
- 15:31Testing for embedded bias
- 21:15Fixing unfair algorithms
- 26:10Ethical model fixes
- 30:10Transparent explainable models
- 33:01Hippocratic oath for data scientists
- 35:08Call to action for ethical ML
related talks
38mSocial Implications of Bias in Machine Learning
47mEthical Machine Learning
38mSocial Implications of Bias in Machine Learning
50mEthical Machine Learning
40mResponsible AI Practices: Fairness in ML
54mEthical Machine Learning
35mBuilding Inclusive Machine Learning Models
42mFairness in Machine Learning: From Theory to Practice