Margaret Mitchell, Senior Research Scientist, Google at MLconf Seattle 2017
summary
Shows that image-caption models learn what humans mention not what’s visually true, because of reporting bias (a 2013 study found “murder” appears ~10× more than “blinking” in web text). Proposes a latent-variable pipeline that separates visual presence and relevance, trains classifiers for both on image-caption data, and marginalizes over them so the model predicts true visual facts even when humans omit them. Demonstrates uncovering implicit gender/race cues and gives a working method that reduces human-label-driven errors.
AI-written summary · about
chapters (4)
- 0:00Start
- 9:10Simple Image Classification
- 10:06Factoring in Reporting Bias: Idea
- 17:18End-to-End Approach
related talks
30mHow to Lie with AI: Understanding Bias, Ethics, and the Hidden Risks in ML
42mKeynote Speaker – Margaret Mitchell
24mDesigning for Transparency in Machine Learning
11mData-centric AI: A new paradigm
38mSocial Implications of Bias in Machine Learning
37mKeynote: Ethical Machine Learning: Creating Fair Models in an Unjust World
31mInclusive Storytelling With Data
41mWhen AI Failures Reflect Our Own Social Biases