Paradoxes in Data Science
summary
Claim: machine learning models mirror cognitive paradoxes when training and test distributions differ, and simulations reveal where they fail. Uses SIR differential equations and agent-based Mesa Streamlit app deployed on AWS with Docker and Plotly; a demo runs a 500-person simulation with ~60 initial infections and a red-box quarantine so policies can be compared. Notes an Imperial-style model with >30 parameters versus the demo’s 6–7, shows Simpson’s paradox flipping correlations and warns about the accuracy, Goldilocks, and unintended-consequence paradoxes.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 1:27Models and simulations
- 4:57Mathematical vs agent models
- 8:10Simulation tools and demos
- 13:04Large language models in simulations
- 15:31Simpson's paradox explained
- 18:19Accuracy paradox in ML
- 20:27Learnability Goldilocks paradox
- 21:31Law of unintended consequences
- 22:55Other paradoxes overview
- 23:37Resources and conclusion
- 24:23Q&A session
related talks
31mUsing Simpson’s Paradox to Discover Interesting Patterns in...
55mEthical Machine Learning
18mThe Importance of Modeling Data Collection
52mBreaking black-box AI
45m"Breaking black-box machine learning"
41mMachine Learning Models in a Dynamic Environment
25mWhy Good Models Fail After Deployment
37mAI/ML Strategies & Techniques for a Post Pandemic World