Interpreting Machine Learning Models: Why and How!
summary
Argues that model interpretability matters for ethics, debugging and trust, and demonstrates R tools: lime and iml to explain random forests on Iris and a bike‑sharing regression, image classifiers via superpixel perturbations, and text sentiment. Claims lime fits a local surrogate by sampling points near an instance and returns a fidelity score but can be unstable; Shapley values apportion the difference from the mean prediction using training data, give theoretically fair attributions but blow up computationally with many features. Shows concrete failures (biased hiring, reduced care hours, snow→wolf).
AI-written summary · about
chapters (12)
- 0:00Intro
- 1:51HOW WOULD YOUR MENTAL MODEL LABEL THIS IMAGE?
- 8:52AUTOMATED REDACTION, TRANSCRIPTION, REPORTING
- 11:34IT IS HUMANS WHO
- 13:37INTERPRETABLE MACHINE LEARNING: WHY?
- 16:10INTERPRETABLE MACHINE LEARNING: HOW?
- 17:06LIME (Tabular Data)
- 22:21LIME (Images)
- 24:49LIME (Text)
- 29:02SHAPLEY VALUES
- 31:06METHODS AND PACKAGES
- 31:37EXTRA READINGS
related talks
15mInterpretable Machine Learning with R
44mInterpretable Machine Learning with Python Examples
28mExplaining Machine Learning Models with LLMs
16mBuilding Trust through Explainable AI
01h 36mOpen the Black Box: an Introduction to Model Interpretability with LIME and SHAP
30mTowards a more transparent AI - Decrypting AI models using LIME
22mInterpreting Machine Learning Models with LIME
46mExplaining a Machine Learning blackbox