Well well ML is not the answer to everything
summary
Compares an expert-system well classifier converted to JS with several ML models trained and evaluated on 800,000 cleaned points and finds the decision-rule experts beat almost every supervised model, notched by an ensemble failure and slow neural nets; expert flowcharts require imputing a missing “staining” feature and still deliver 81–85% accuracy with high specificity, exposing how opaque models and marginal accuracy gains can’t answer ethical deployment questions like false negatives, adherence and who decides acceptable harm.
AI-written summary · about
chapters (21)
- 0:00Introduction
- 0:30Groundwater arsenic pollution
- 1:23The model
- 2:04Data
- 3:32A complete tangent
- 4:20How to make a predictive model
- 6:12Decision boundary
- 7:37Making evaluation
- 8:12Accuracy
- 9:11Confusion Matrix
- 9:52How good is good enough
- 10:26Magic model
- 10:58Magic model results
- 12:14The barrier
- 13:40The colonial mindset
- 16:30Cats from dogs
- 17:14How much did my ML models outperform
- 18:32Bridging the gap
- 20:41Interdisciplinary projects
- 21:19Survivorship bias
- 22:04Thank you
related talks
12mOn the Use of ML for Blackbox System Performance Prediction
25mThe Uncanny Valley of ML
35mML: Beyond confusion matrix, model performance analysis
09mAI/ML Won't Save you
39mAll of Our ML Ideas Are Bad (and We Should Feel Bad)
44mAutomated Machine Learning with ML.NET
44mAI Engineering Anti-patterns
44mThe ML Mindset for Managers