Product fairness testing for Machine Learning developers
summary
A practical workflow for locating and removing biased prompts from a generative-text corpus: load the WikiDialog sample (1,115) in Colab, consult the dataset’s data card, run a match of content against a sensitive terms dataset (590 items) and co-occurrence rules, use the Google Cloud ML sentiment classifier to tag adjectives, and surface coincidences. The pipeline finds 15 problematic interactions, and a simple mitigation is to delete those instances while preserving legitimate examples.
AI-written summary · about
chapters (14)
- 0:00Introduction to AI Fairness
- 0:42Workshop Overview
- 2:09Google's AI Principles
- 4:10Defining Algorithmic Bias
- 6:28Case Study: Wiki Dialogue
- 9:54Fairness Testing Framework
- 13:35Exploring the Data Set
- 15:06Identifying Unfair Bias
- 18:40Extracting Problematic Data
- 23:47Refining Search for Bias
- 27:51Evaluation and Analysis
- 30:35Mitigation Strategies
- 31:01Limitations and Constraints
- 32:33Workshop Summary
related talks
01h 30mFairness Tales: How To Measure And Mitigate Unfair Bias in Machine Learning Models
47mEthical Machine Learning
40mResponsible AI Practices: Fairness in ML
36mMachine Learning Fairness: Lessons Learned
37mKeynote: Ethical Machine Learning: Creating Fair Models in an Unjust World
39mBeyond Bias: Crafting Fairness in Pythonic Machine Learning Systems
31mFlagging *and* Fixing Bias in ML
28mBreaking the Chains of Oppressive Software