What We Expect from XAI - A scientist’s experience
summary
Project applies SHAP to remote sensing models for urban resilience, builds maps and a dashboard that surfaces top contributors but requires heavy curation. Shows SHAP values are not causal, demonstrates that background dataset selection changes results, and reports high compute cost with approximation-sensitive implementations. Explains misinterpretations by domain users, how correlated features redistribute importance, and that explanations only reflect model behaviour not ground truth — useful if ready to fix data and feature design.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 1:10Why explainability in AI matters
- 2:55Performance vs interpretability
- 5:27Good explanations perspectives
- 7:35Explainability, interpretability, trust
- 9:19XAI methods categories
- 11:26XAI in environmental sciences
- 13:55SHAP popularity and limits
- 14:38Mapping risks in Southern Italy
- 17:26How SHAP works
- 20:58Visualizing SHAP results
- 22:05SHAP limitations
- 24:54Implementation challenges
- 27:03Communicating XAI results
- 29:17Conclusions and key takeaways from the talk
related talks
36mSHAP beyond the standard graphics: co-design of ML-models in earth sciences
31mBeyond the Black Box: Interpreting ML Models with SHAP
33mXAI meets Natural Language Processing
30mTrustworthy AI - opening up the black box
39mThe Search for Transparency and Accountability in the Age of...
35mSHAPtivating Insights: unravelling blackbox AI models
26mRevealing the Unseen - Leveraging XAI for Deeper Data Insights
27mInterpretability and explainability in machine learning