“Why Should I Trust You?” - Debugging black-box text classifiers
summary
Claims that eli5 and LIME produce faithful local explanations of black‑box text classifiers by training a simple surrogate (CountVectorizer + logistic regression) on thousands of perturbed samples (about 5000 perturbations), and shows concrete results on 20 Newsgroups where explanations reach ~90% local fidelity but sometimes fail — e.g. revealing a document-length leakage and domain-name features, and demonstrating a practical fix (add length feature) and the limits with character models.
AI-written summary · about
chapters (10)
- 0:00Model trustworthiness
- 1:48Validation sets insufficient
- 2:50LIME algorithm overview
- 5:23How LIME generates explanations
- 7:30Debugging a text classifier
- 13:56Eli5 for visualization
- 16:02Limitations and failure cases
- 19:19Key takeaways
- 20:42Speaker and resources
- 21:34Q&A
related talks
16mBuilding Trust through Explainable AI
30mTowards a more transparent AI - Decrypting AI models using LIME
39mKeynote 3: How did you know? Explaining Black Box Model Predictions in Python
34mInterpreting Machine Learning Models: Why and How!
46mExplaining a Machine Learning blackbox
17mChallenges Towards Production-Ready Explainable Machine Learning
01h 04mCan you trust your (large language) model?
50mDehumanizing Agents: Why Explainability is Crucial in the LLM Era