Towards a more transparent AI - Decrypting AI models using LIME
summary
Demonstrates how LIME image explainer creates sparse, local linear explanations by perturbing input and fitting a local surrogate linear model to reveal which pixels drive predictions. Runs demos on dog vs cat classifiers and vision models using ~100–500 perturbation samples, finds many false positives driven by background or watermarks, and shows simple code steps to generate perturbed samples and fit explanations. Claims LIME helps find mislabeled data and spurious features but breaks when proximity assumptions fail or when explanations rely on too few samples.
AI-written summary · about
chapters (7)
- 0:00Introduction to AI transparency
- 3:02Why AI models need explanation
- 8:28Understanding LIME for model interpretation
- 15:00Practical examples of LIME
- 22:12Building trust with transparent AI
- 26:12Exploring LIME research paper
- 28:49Alternatives and future of explainable AI
related talks
33mExplainable AI in the LIME-light
16mBuilding Trust through Explainable AI
22mInterpreting Machine Learning Models with LIME
46mAI for High-Stakes Decision Making
34mInterpreting Machine Learning Models: Why and How!
45mExplainable Machine Learning Explained
44mInterpretable Machine Learning with Python Examples
23m“Why Should I Trust You?” - Debugging black-box text classifiers