Towards Inclusive AI: Understanding bias in ML models
summary
Machine learning pipelines amplify societal bias despite safety layers and the talk tests that with a Llama 2 case study and a manual audit of 20,000 model outputs labelled across sizes. The pipeline steps are listed—web scraping, prefiltering, pretraining, diversity fine‑tuning, RLHF—but results show safer models trade explicit slurs for refusal-to-answer bias and unequal service: larger models refuse or give harmful rationales more often for ethnically marked names, so last‑step mitigation reduces visible toxicity without removing underlying stereotypes.
AI-written summary · about
chapters (8)
- 0:00Introduction to speakers
- 1:30What is AI bias
- 11:30Sources of bias in ML
- 17:07Bias in large language models
- 23:36Industry safety pipeline
- 31:07Testing safety safeguards
- 39:59Why safety fails
- 40:41Solutions and future work
related talks
35mBuilding Inclusive Machine Learning Models
50mEthical Machine Learning
23mRemoving Unfair Bias in the Machine Learning Pipeline
38mSocial Implications of Bias in Machine Learning
38mSocial Implications of Bias in Machine Learning
56mAn AI with an Agenda: How Our Cognitive Biases Leak Into Machine Learning
31mFlagging *and* Fixing Bias in ML
35mZero Bias AI