AI-404: Explanation Not Found
summary
The talk examines explainability of modern AI systems, contrasting their good performance on tasks like image generation with their poor ability to justify specific outputs (e.g., why a generated image is a cat). It discusses how advances — faster models, deeper neural networks and black-box architectures used in deep learning for image recognition and NLP — have made obtaining human-understandable explanations difficult and asks whether AI can still be explained or has become beyond our control.
AI-written summary · about
chapters (24)
- 0:00Introduction
- 1:04Data scientist role
- 3:30Banking decision example
- 5:16AI definition and math
- 9:52Simple artificial neuron
- 10:14What is XAI?
- 13:26Ethical and legal issues
- 15:35Strategic challenges
- 18:25XAI technical methods
- 20:17XAI scope and dependencies
- 21:22LIME method
- 24:30SHAP method
- 27:57Global methods and PDP
- 30:02Deep Learning explainability
- 31:27Deep Learning feature visualization
- 33:34Interpretable feature extraction
- 36:01XAI research history
- 37:48Client interface Act
- 39:15Generative AI limits
- 40:18Interpretability open source
- 41:21XAI evolution predictions
- 42:05Good XAI practices
- 43:10Risks EU Act compliance
- 43:52Conclusion
related talks
45mAI-404: explanation not found
32mOn the Importance of Explainability
41mThe Explainability Problem: Towards Understanding AI
35mWhy Do You Need an Explainable Ai?
43mBring in the Lawyers: Explainable AI Driven Decision-making
24mWhy and How to Build Explainability into your ML Workflow
19mMoving Forward with Explainable AI: Adoption Rate & Essential Dimensions for a Reliable Use
41mIf We Want AI to be Interpretable, We Need to Measure Interpretability