Ethical Machine Learning
summary
Predictive policing models, COMPAS recidivism scores and Amazon hiring models actually misfire; hands-on demo uses SageMaker Clarify to run a processing job that outputs bias reports, finding class imbalance 0.557 and DPL 0.18 on raw stop‑and‑search data. The workflow shows exact checkpoints: dataset wrangling, eight bias metrics, model explainability reports, and continuous monitoring for drift. Practical takeaway: train on data that reflects the deployment population, set acceptable false‑positive/negative thresholds, and log predictions so decisions remain auditable.
AI-written summary · about
chapters (38)
- 0:00Introduction and use cases
- 2:08Ethical dilemmas in AI
- 5:07Speaker background
- 7:51What is AI?
- 9:07Pervasive machine learning
- 11:24How ML systems work
- 11:52Traditional vs ML development
- 12:42The ML development lifecycle
- 13:37Garbage in, garbage out
- 15:37Defining ethical AI
- 16:55The ethics of building
- 18:30Bias in AI systems
- 19:23How bias surfaces
- 19:58Lessons from failures
- 21:44Bias in action
- 23:24Human decision history
- 24:40Impact of unfair decisions
- 26:23Mitigating bias locations
- 28:22People-focused bias
- 29:43Strategic risk management
- 31:49Data-focused risk
- 34:08Algorithmic and model bias
- 35:30Action-based ethics
- 36:29Mitigation strategies
- 37:30Explainable AI
- 38:13Incorporating checks
- 39:19Automation with AWS
- 40:52Features of Clarify
- 41:37Incorporating Clarify
- 42:34SageMaker demo
- 43:28Data inspection
- 43:56Bias metrics overview
- 45:15Demo: Pre-training report
- 45:58Interpreting results
- 49:20Legal responsibilities
- 52:01Call to action
- 52:40Further learning
- 53:09Final Q&A
related talks
50mEthical Machine Learning
54mEthical Machine Learning
47mEthical Machine Learning
41mMachine Learning Model Fairness and Interpretability with AWS
37mKeynote: Ethical Machine Learning: Creating Fair Models in an Unjust World
38mSocial Implications of Bias in Machine Learning
23mRemoving Unfair Bias in the Machine Learning Pipeline
33mBuild trustworthy ML detection and monitor bias