![Visualization of CNN
Feature Evolution during Training
Epoch
=[1, 2, 5, 10, 20, 30, 40, 64]](https://image.slidesharecdn.com/visualizaingandunderstandingconvolutionalnetworks-190711043511/85/Visualizaing-and-understanding-convolutional-networks-21-320.jpg)
Visualizaing and understanding convolutional networks
- 1. Comprehension of deep-learning - Visualizing and Understanding Convolutional Networks 17.01.06 You Sung Min Zeiler, Matthew D., and Rob Fergus. "Visualizing and understanding convolutional networks.“ European Conference on Computer Vision. Springer International Publishing, 2014. Paper review
- 2. 1. Review of Deep learning (Convolutional Neural Network) 2. Visualization of CNN 3. Feature generalization (Transfer learning) Contents
- 3. Structure of Neural Networks A simple model to emulate a single neuron This model produces a binary output Review of Deep learning = 𝟎 𝒊𝒇 𝒋 𝝎𝒋 𝒙𝒋 ≤ 𝑻 𝟏 𝒊𝒇 𝒋 𝝎𝒋 𝒙𝒋 > 𝑻 𝝎 𝟏 𝝎 𝟐 𝝎 𝟑 𝒋 𝝎𝒋 𝒙𝒋Inputs Threshold T Perceptron (1950) Neuron
- 4. Review of Deep learning Multilayer Perceptron (MLP) A network model consists of perceptrons This model produces vectorized outputs
- 5. Multilayer Perceptron (MLP) Review of Deep learning Handwritten digit with 28 by 28 pixel image Binary Input (Intensity of a pixel) 28 28 Input (784) Desired output for “5” 𝒚(𝒙) = 𝟎, 𝟎, 𝟎, 𝟎, 𝟏, 𝟎, 𝟎, 𝟎, 𝟎 𝑻
- 6. Convolutional Neural Network Convolution layer Subsampling (Pooling) layer Rectified Linear Unit(ReLU) Review of Deep learning Feature Extractor Classifier
- 7. Convolutional Neural Network Review of Deep learning
- 8. Convolutional Neural Network Review of Deep learning y = max(x,0)
- 9. Convolutional Neural Network Review of Deep learning Feature map
- 10. Convolutional Neural Network Review of Deep learning Feature Extractor Classifier Feature map
- 11. Visualization of CNN Deconvnet (Deconvolutional Network) Mapping the activations back to the input pixel space What input pattern caused activation in the feature map → Reconstruct input space with feature map Feature map
- 12. Visualization of CNN Stacked-Autoencoder (SAE) Generative model with RBM Produce same output with the input
- 13. Visualization of CNN Deconvnet (Deconvolutional Network) Deconvnet CNN Feature maps Normalization Unpooling Rectify Deconvolution Input Image
- 14. Visualization of CNN Deconvnet (Deconvolutional Network) Deconvnet CNN
- 15. Visualization of CNN Architecture of network CNN with 8 layers (5 as convolution, 3 for MLP) Trained with ImageNet 2012 - 1.3 million images with 1000 classes Train took around 12 days with GTX 580
- 16. Visualization of CNN Visualization of feature map Layer 2 - Corner, Edge Layer 3 - Texture, Text Reconstructed Image Corresponding input images
- 17. Visualization of CNN Visualization of feature map Layer 4 - Object Layer 5 - Object with pose variation
- 18. Visualization of CNN Visualization of feature map The network is trained discriminatively, those features maps (strong activations) shows which part of the input image are discriminative
- 19. Visualization of CNN Effect of occlusion Changes in output and feature map with different portions of gray square
- 20. Visualization of CNN Visualization of feature map Yosinski, Jason, et al. "Understanding neural networks through deep visualization."
- 21. Visualization of CNN Feature Evolution during Training Epoch =[1, 2, 5, 10, 20, 30, 40, 64]
- 23. Feature generalization Caltech 101 classification accuracy
- 24. Feature generalization Caltech 256 classification accuracy
- 25. Feature generalization PASCAL 2012 classification accuracy Due to the inequality of the dataset type
- 26. References Image Source from https://deeplearning4j.org/convolutionalnets Zeiler, Matthew D., and Rob Fergus. "Visualizing and understanding convolutional networks.“ European Conference on Computer Vision, Springer International Publishing, 2014. Jia-Bin Huang, “Lecture 29 Convolutional Neural Networks”, Computer Vision Spring 2015 Yosinski, Jason, et al. "Understanding neural networks through deep visualization."
Editor's Notes
- #27: 13층의 컨볼루션 신경망의 값을 산출하기 위해선 약 300억 번의 연산수 필요