Bringing life and motion to AI Explainability
summary
Project predicts progression to severe chronic kidney disease using SHAP values to track feature importance at training checkpoints; SHAP is always evaluated on unseen test folds. Pipeline feeds per-checkpoint SHAP activations into an LSTM + autoencoder to produce a 2D embedded conformational space that clusters ~10 algorithms by learning trajectory. Result: gradient-boosted models show stable top features, Gaussian Naive Bayes uniquely flags procedure code 83970, decision tree and Bernoulli NB behave similarly at very low samples, and animated charts expose instability humans can adjudicate.
AI-written summary · about
chapters (12)
- 0:00Introduction to AI explainability
- 1:10Clinical AI project background
- 2:32Limitations of traditional explainability
- 4:41Trust and model selection criteria
- 5:45Tracking model learning stability
- 7:11Visualizing model training evolution
- 9:21From black box to glass box
- 10:03Dynamic explainability animations
- 13:13Cross-comparing algorithms
- 17:02Embedded model comparison
- 22:00Enhancing AI safety in healthcare
- 22:47Q&A session
related talks
31mBeyond the Black Box: Interpreting ML Models with SHAP
35mSHAPtivating Insights: unravelling blackbox AI models
32mSHAP and Beyond
10mI Think I Get Your Point, AI! The Illusion of Explanatory Depth in Explainable AI
27mInterpretability and explainability in machine learning
39mExplainable AI as a Conventional Data Analysis Tool
35mTaking a peek under the hood - Interpreting black box model
28mManaging AI Risk: Beyond Explainability