Let's Talk about AI, ML and Bias
summary
Explains that voice recognition fails on Scottish accents and women's voices; a medical app cannot track women's cycles affecting 50% of the population; facial recognition uses a training set that is 77% white male and mislabels Black women; COMPAS risk scores perform no better than a random person $1; Amazon's hiring model downgrades female candidates; a hospital triage model ignores asthma because the training data omitted it. Argues for diverse teams, data audits, alternate models, monitoring and explainability.
AI-written summary · about
chapters (14)
- 0:00Let's Talk about AI, ML and Bias
- 3:26Professional Athlete
- 10:18And what about manipulation?
- 11:25The best intentions can be undone by bias.
- 19:47Areas of expertise
- 21:26"The real safety question, if you want to call it that is that if we give these systems biased data, they will be biased," - John Giannandrea, Google's Al Chief
- 23:58Bias through interaction Learned/taught bias
- 24:29Emergent bias Bias bubbles
- 26:07Inherent data-driven bias Flawed data sets
- 30:27Consider data and algorithmic remediation.
- 31:53Testing can't stop once a system is unleashed.
- 33:49Beware the data land grab.
- 34:39Al needs to show its work.
- 37:44I'm biased. You're biased. The data is biased.
related talks
38mSocial Implications of Bias in Machine Learning
27mCrouching Harm, Hidden Bias—Addressing issues of fairness....
47mEthics & AI: Identifying & preventing bias in predictive models
54mEthical Machine Learning
22mGender Bias in Artificial Intelligence
21mFairness and bias in AI
26m(Un)ethical Artificial Intelligence: How to Keep the AI Fair for Everyone
38mSocial Implications of Bias in Machine Learning