Explainable Machine Learning Explained
summary
Claims: convolutional models reach ~90% on image tasks but learn spurious cues (husky/wolf uses snow), and FDA needs explainability for MedTech. Tools shown: decision trees and linear models for white‑box baselines, Class Activation Map for CNN heatmaps, and model‑agnostic methods like LIME and SHAP and one-pixel adversarial attacks to probe robustness. Demos live on GitHub (Titanic, skin cancer, VGG16 vs Inception) and illustrate that bad data, not model choice, often breaks performance.
AI-written summary · about
chapters (15)
- 0:00Introduction to explainable machine learning
- 1:39Why explainability matters
- 3:24AI rise and explainability
- 5:55Reasons for explainable AI
- 7:59Types of AI explanations
- 9:22Feature relevance and debugging
- 13:19White vs black box ML
- 17:33Insights via explainability
- 21:26Black box limitations
- 24:57Model-agnostic explainability
- 26:45Libraries for explainable AI
- 30:21Deep learning challenges
- 35:21Adversarial attacks and explainability
- 37:53Summary: explainable AI
- 38:35Q&A on AI trade-offs
related talks
46mExplaining a Machine Learning blackbox
24mWhy and How to Build Explainability into your ML Workflow
16mBuilding Trust through Explainable AI
15mIntroduction to Explainable AI
37mExplaining model explainability
17mChallenges Towards Production-Ready Explainable Machine Learning
29mTowards Explainable Machine Learning
35mDeep Learning Model Interpretability for Computer Vision based Models