Deep Learning Model Interpretability for Computer Vision based Models
summary
Shows concrete methods and outcomes: uses shap and lime and gradient heat maps like grad cam to localize features, runs PCA/t‑SNE on embeddings for clustering, and uses Keras‑wiz for layer inspection. Reports a surprising training artifact — loss going negative and a stuck ~0.008 loss with poor inference — and fixes it by data cleaning and one‑shot/heuristic sampling. Built detectors for fire extinguishers, signs and medical lesions; visualizations help find spurious signals (lights, backgrounds) and guide fixes that sometimes succeed and sometimes don’t.
AI-written summary · about
chapters (19)
- 0:00Introduction
- 1:50Deep learning for computer vision
- 3:39Model interpretability matters
- 5:06Explaining models to business
- 6:32Debugging with loss functions
- 8:59Explainable AI frameworks
- 10:24SHAP and LIME insights
- 13:39Gradient-based visualization
- 15:47Visualization tools
- 17:53Heuristic learning for detection
- 20:03Metrics to business KPIs
- 21:45Healthcare diagnostics
- 23:10Training and debugging neural nets
- 26:10Visualization reliability
- 28:41Pattern recognition case study
- 29:03Cost of explainability
- 30:49Convincing clinicians
- 33:21GPU costs in training
- 34:49Conclusion and Q&A
related talks
45mExplainable Machine Learning Explained
34mInterpreting Machine Learning Models: Why and How!
27mUnderstanding Deep Neural Networks
30mTowards a more transparent AI - Decrypting AI models using LIME
16mBuilding Trust through Explainable AI
35mDeconstructing Deep Learning
35mVisual Diagnostics for More Effective Machine Learning
01h 01mA practical guide to deep learning