Open Source Fairness
summary
Shows concrete failures and concrete fixes: algorithms using proxy variables like healthcare spending reproduce racial gaps (Science study) and credit-screening asymmetries (Apple Card) rather than solving them. Provides implemented interventions — aif 360 bundles state of the art algorithms for pre/in/post processing, iqtas and fairml audit models and feature influence, and ml fairness gym simulates feedback loops so teams can test whether an intervention actually reduces disparity over time.
AI-written summary · about
chapters (16)
- 0:00Open source fairness intro
- 1:06Fairness in AI unsolved
- 3:11AI discrimination examples
- 6:22Data proxies and incompleteness
- 9:31Labeling biases consequences
- 12:40ML amplifies bias
- 15:04Accuracy versus fairness
- 16:48Why removing attributes doesn't ensure fairness
- 17:52Defining fairness in ML
- 21:45Fairness beyond anti-discrimination
- 23:58Fairness interventions pipeline
- 26:25Pre-processing fairness
- 27:28In-processing loss functions
- 28:09Post-processing discrimination
- 28:52Open-source fairness tools
- 34:07Summary & Q&A
related talks
53mIt’s Not Fair! Detecting Algorithmic Bias with Open Source Tools
42mFairness in Machine Learning: From Theory to Practice
40mEthical AI: Addressing Bias & Algorithmic Fairness
40mQuestions and Practices to Make Algorithmic Decision Making More Fair
35mBuilding Inclusive Machine Learning Models
37mKeynote: Ethical Machine Learning: Creating Fair Models in an Unjust World
34mInterpretability and Algorithmic Fairness
23mRemoving Unfair Bias in the Machine Learning Pipeline