Understanding Deep Neural Networks
summary
Clear claims and concrete tools: model attributions use captum on a pretrained ResNet50 with top‑3 outputs, showing gradients (saliency, input×gradient, guided backprop, guided grad‑cam) and perturbation (occlusion with 60x60 sliding window). A demonstrated failure mode is a horse classifier relying on a watermark, not the animal. Recommendation is to run both backprop and occlusion, generate per‑epoch attribution images, and consider SHOP (sharply added explanations) for principled feature attributions.
AI-written summary · about
chapters (10)
- 0:00Deep neural networks intro
- 2:13Why explainability matters
- 5:42Explainable AI methods
- 9:13Model-agnostic vs model-specific
- 10:56Explaining neural networks
- 13:12Model visualization challenges
- 14:36Live Captum demo
- 20:50Gradient vs perturbation methods
- 22:38References and further reading
- 23:59Q&A
related talks
08mModel Interpretability with Captum
27mDemystifying Deep Neural Networks
34mNeural networks: insane in the membrane
35mDeconstructing Deep Learning
01h 02mTutorial on Deep Learning I
35mDeep Learning Model Interpretability for Computer Vision based Models
54mIntroduction to Understanding Deep Learning
47mInterpreting Deep Neural Networks (DNNs)