Keynote address
summary
Team scrapes 866,000 web bios from Common Crawl and labels them to 33 job categories, finds the corpus is 56% male/44% female, trains bag‑of‑words, word‑embedding, and attention DNN classifiers, and shows swapping he/she flips predicted occupations (surgeon→nurse, teacher→pastor) with a measurable gender accuracy gap that compounds hiring imbalances. Mitigation comes from penalizing name–accuracy correlations during training, which shrinks gender and racial gaps and earned best‑paper awards.
AI-written summary · about
chapters (13)
- 0:00Introduction of keynote speaker
- 2:12Unconventional path to data science
- 9:20Early career in math and physics
- 13:30Transition to Microsoft Research
- 21:52Research on fairness in machine learning
- 24:23Gender bias in occupation prediction
- 34:22Why I joined UC Berkeley
- 38:25Vision for Computing, Data Science, and Society
- 40:33Q&A on interdisciplinary collaboration
- 44:10Q&A on funding and humanities
- 47:12Q&A on outreach and resources
- 49:39Q&A on business school integration
- 51:32Closing remarks and thanks
related talks
27mDesigning More Inclusive AI with More Inclusive Workflows
35mBias in NLP 101
31mCombining behavioral research & NLP to combat gender biases
31mCombining behavioral research and NLP to combat gender biases
28mKeynote, Why and how to care about ethics in ML
28mBreaking the Chains of Oppressive Software
28mBias in NLP
22mGender Bias in Artificial Intelligence