Responsible AI Practices: Fairness in ML
summary
Clear claims and concrete tools are described so watch if you want methods not platitudes. Shows a measured claim that pre trained embeddings encode stereotypes with a demonstrated effect on sentiment scores, gives the word-embedding gender and race bias experiment and downstream smart‑reply example, documents problematic outputs from the Perspective API toxicity scores (0.8 0.5 0.46) and how adding non‑toxic examples reduces those errors, explains incompatible fairness definitions with a loan demo (profit example ~$32,400), and demos Facets and What‑If tools that are usable now.
AI-written summary · about
chapters (12)
- 0:00Responsible AI practices
- 1:51Google's AI principles
- 2:55Fair ML challenges
- 6:45Bias in word embeddings
- 12:09Testing fairness contexts
- 16:06Defining fairness
- 21:28Tools: Facets & What-If
- 26:46Inclusive ML resources
- 27:31Q&A 1
- 30:05Q&A 2
- 33:57Q&A 3
- 37:57Q&A 4
related talks
37mKeynote: Ethical Machine Learning: Creating Fair Models in an Unjust World
40mEthical AI: Addressing Bias & Algorithmic Fairness
26mBuilding ethical AI: Transparency, bias mitigation, and trust
36mMachine Learning Fairness: Lessons Learned
39mResponsible AI
41mUsing AI to Build Fair and Equitable Workplaces
23mRemoving Unfair Bias in the Machine Learning Pipeline
47mEthical Machine Learning