SLIDE 1
Interpretability of Machine Learning for Computer Vision
Xinshuo Weng*
*Most slides borrowed from CVPR 2018 tutorial and Stanford
Type of interpretability methods
Interpretability of Machine Learning for Computer Vision Xinshuo - - PowerPoint PPT Presentation
Interpretability of Machine Learning for Computer Vision Xinshuo - - PowerPoint PPT Presentation
Interpretability of Machine Learning for Computer Vision Xinshuo Weng* *Most slides borrowed from CVPR 2018 tutorial and Stanford Type of interpretability methods Type of interpretability methods Type of interpretability methods Type of
SLIDE 2
SLIDE 3
Type of interpretability methods
SLIDE 4
Type of interpretability methods
SLIDE 5
Type of interpretability methods
SLIDE 6
Understanding Models at Different Granularity
- What is a unit doing?
- What are all the units are doing?
- How units are relevant to prediction? Understanding the explainable model.
SLIDE 7
- Visualize the filter
What is a unit doing? - Visualize the individual unit
Krizhevsky et al, “ImageNet Classification with Deep Convolutional Neural Networks”, NIPS, 2012.
SLIDE 8
- Visualize the filter
- Visualize the activation
What is a unit doing? - Visualize the individual unit
Yosinski et al, “Understanding Neural Networks Through Deep Visualization”, ICML, 2015.
SLIDE 9
- Visualize the filter
- Visualize the activation
- Visualize the corresponding images
- Top activated images (NN)
What is a unit doing? - Visualize the individual unit
Yosinski et al, “Understanding Neural Networks Through Deep Visualization”, ICML, 2015.
SLIDE 10
- Visualize the filter
- Visualize the activation
- Visualize the corresponding images
- Top activated images (NN)
- Deconvolution
What is a unit doing? - Visualize the individual unit
Zeiler and Fergus, “Visualizing and Understanding Convolutional Networks”, ECCV, 2014.
SLIDE 11
- Visualize the filter
- Visualize the activation
- Visualize the corresponding images
- Top activated images (NN)
- Deconvolution
- Back-propagation
What is a unit doing? - Visualize the individual unit
SLIDE 12
- Visualize the filter
- Visualize the activation
- Visualize the corresponding images
- Top activated images (NN)
- Deconvolution
- Back-propagation
What is a unit doing? - Visualize the individual unit
Erhan et al, “Visualizing Higher-Layer Features of a Deep Network”, University of Montreal, 2009.
SLIDE 13
- Visualize the filter
- Visualize the activation
- Visualize the corresponding images
- Top activated images (NN)
- Deconvolution
- Back-propagation
What is a unit doing? - Visualize the individual unit
Springenberg et al, “Striving for Simplicity: the All Convolutional Net”, ICLR 2015.
SLIDE 14
- Visualize the features
- Dimensionality reduction: t-SNE
What are all the units doing?
PCA t-SNE
Maaten and Hinton, “Visualizing Data using t-SNE”, JMLR, 2008.
SLIDE 15
- Visualize the features
- Dimensionality reduction: T-SNE
- Visualize the corresponding images
- Top activated image (NN)
What are all the units doing?
Krizhevsky et al, “ImageNet Classification with Deep Convolutional Neural Networks”, NIPS, 2012.
SLIDE 16
- Visualize the features
- Dimensionality reduction: T-SNE
- Visualize the corresponding images
- Top activated image (NN)
- Back-propagation
What are all the units doing?
Ulyanov et al, “Deep Image Prior”, CVPR 2018.
SLIDE 17
- Visualize the features
- Dimensionality reduction: T-SNE
- Visualize the corresponding images
- Top activated image (NN)
- Back-propagation
What are all the units doing?
Ulyanov et al, “Deep Image Prior”, CVPR 2018.
SLIDE 18
- Visualize the features
- Dimensionality reduction: T-SNE
- Visualize the corresponding images
- Top activated image (NN)
- Back-propagation
- Image Synthesis
What are all the units doing?
Dosovitskiy And Brox, “Generating Images with Perceptual Similarity Metrics based on Deep Networks”, NIPS 2016.
SLIDE 19
- Visualize the features
- Dimensionality reduction: T-SNE
- Visualize the corresponding images
- Top activated image (NN)
- Back-propagation
- Image Synthesis
What are all the units doing?
Dosovitskiy And Brox, “Generating Images with Perceptual Similarity Metrics based on Deep Networks”, NIPS 2016.
SLIDE 20
- From qualitative to quantitative analysis: Network Dissection
What are all the units doing?
Bau et al, “Network Dissection: Quantifying Interpretability of Deep Visual Representations”, CVPR 2017.
SLIDE 21
- From qualitative to quantitative analysis: Network Dissection
What are all the units doing?
Bau et al, “Network Dissection: Quantifying Interpretability of Deep Visual Representations”, CVPR 2017.
SLIDE 22
- Ablation study: occlusion effect
How units are relevant to prediction? Understanding the explainable model
Zhou et al, “Revisiting the Importance of Individual Units in CNNs via Ablation”, arXiv 2018.
SLIDE 23
- Ablation study: occlusion effect
How units are relevant to prediction? Understanding the explainable model
Zhou et al, “Revisiting the Importance of Individual Units in CNNs via Ablation”, arXiv 2018.
SLIDE 24
- Ablation study: occlusion effect
How units are relevant to prediction? Understanding the explainable model
Zhou et al, “Revisiting the Importance of Individual Units in CNNs via Ablation”, arXiv 2018.
SLIDE 25
- Ablation study: occlusion effect
How units are relevant to prediction? Understanding the explainable model
Zhou et al, “Revisiting the Importance of Individual Units in CNNs via Ablation”, arXiv 2018.
SLIDE 26
- Ablation study: occlusion effect
- For sequential modeling: true for different model configuration
- Range of context (memory) is limited – 200 tokens
- Order matters in nearby context (not long-range context) – 50 tokens
How units are relevant to prediction? Understanding the explainable model
Khandelwal, et al, “Sharp Nearby, Fuzzy Far Away: How Neural Language Models Use Context”, ACL 2018.
SLIDE 27
- How do we improve the mode based on the interpretability?
Conclusion