Diversity and Inclusion for Algorithmic Bias Considerations
summary
The talk details concrete failures and fixes: the Gender Shades study shows industry training sets skew toward light faces and that using a balanced dataset improves accuracy for minorities and for everyone; ProPublica’s analysis of COMPAS finds unequal false positive/negative error types across races; simple sensors like soap dispensers fail on dark skin; big tech reports ~3.5% Black workers vs US census 13.4% (Apple ~8.5%). Standards work named includes P7003 algorithmic bias considerations, IEEE P7000/P7010 and ISO/IEC JTC1 SC42 WG3 on trustworthiness.
AI-written summary · about
chapters (19)
- 0:00Introduction
- 2:08Algorithmic Bias
- 4:05Gender Shades
- 6:02Compass
- 9:11Failure of Diversity
- 18:08Diversity in Silicon Valley
- 20:54AI Principles
- 23:09AI Principles Analysis
- 24:00AI Race
- 25:22AI Policy Bodies
- 27:36Governance Frameworks
- 33:59Standards Developments
- 37:10AI Ethics Global Initiative
- 37:54Algorithmic Bias Considerations
- 39:01What is Bias
- 39:47Causes of Bias
- 42:16Cultural Aspects
- 43:47Bias Issues
- 44:07AI Ethics Diversity
related talks
27mCrouching Harm, Hidden Bias—Addressing issues of fairness....
45mBias in Artificial Intelligence
41mMeasures and Mismeasures of Algorithmic Fairness
38mSocial Implications of Bias in Machine Learning
41mAvoiding bias in software design
47mEthics & AI: Identifying & preventing bias in predictive models
38mSocial Implications of Bias in Machine Learning
21mFairness and bias in AI