Explainable AI in the LIME-light
summary
LIME claims to explain individual predictions by creating perturbations around one instance and fitting a local linear surrogate so you see which features drove a single decision. The demo uses the Python package lime (pip install lime) with a random-forest on NBA tabular data, a text classifier on a PSV news article, a neural net on an image (bike vs unicycle) and an open-source lime_for_time_series ECG extension. Outputs highlight contributing features or time segments, run faster than kernel SHAP, but perturbations can produce invalid synthetic points.
AI-written summary · about
chapters (22)
- 0:00Introduction
- 0:43What is explainable AI
- 2:29Reasons for explainability
- 4:40Global vs local explainability
- 6:04Introducing LIME
- 7:09How LIME works
- 8:54Installing LIME in Python
- 10:00Explaining tabular data
- 11:23Explaining text data
- 12:48Explaining image data
- 15:37LIME for time series
- 18:58Why use LIME
- 20:03Q&A on ranking & regression
- 21:13When LIME gives valid explanations
- 23:49LIME vs SHAP speed
- 24:53Explaining LIME to experts
- 26:06Validating medical perturbations
- 27:57Exploring local points
- 29:10Does LIME estimate accuracy?
- 30:59LIME vs saliency maps
- 31:43Handling LIME errors
- 32:28Q&A
related talks
30mTowards a more transparent AI - Decrypting AI models using LIME
16mBuilding Trust through Explainable AI
44mInterpretable Machine Learning with Python Examples
22mInterpreting Machine Learning Models with LIME
24mOpen and Understandable: Designing Explainable AI for Everyone
46mAI for High-Stakes Decision Making
39mKeynote 3: How did you know? Explaining Black Box Model Predictions in Python
34mInterpreting Machine Learning Models: Why and How!