[深度学习论文笔记][Visualizing] Understanding Neural Networks Through Deep Visualization
2016-10-31 08:27
471 查看
Yosinski, Jason, et al. “Understanding neural networks through deep visualization.” arXiv preprint arXiv:1506.06579 (2015). (Citations: 65).
1 Optimization For Any Arbitary Neuron
2 Three Forms of Regularizations
This paper proposed some more explicit forms of regularization. Different hyperparameter combinations of these regularizations produce different styles of reconizable images.
.2.1 Clipping Pixels with Small Contribution
Contribution can be measured as
Where Ã(d, i, j) is the activation with a certain pixel of X be set to zero.
However, this approach is slow since it needs to forward pass for every pixel. So we approximate Ã(d, i, j) by
So
Finally, we sum across channels to find pixels with small distribution in either direction, positive or negative, and set those pixels to zero by thresholding.
2.2 Blur X a Bit
Producing images via gradient ascent tends to produce examples with high frequency information. While these images cause high activations, they are neither realistic nor interpretable. As a result, we blur the image with Gaussian kernel every several steps
due to the computational effectiveness.
2.3 Take Any Pixel With Small Norm to Zero
CNN tends to give X non-zero everywhere to contribute in whatever small way they can to ultimately raise the chosen unit’s activation. We instead show only the main object, letting other regions be exactly zero by thresholding if they are not needed.
3 References
[1]. Deep Visualization Toolbox. https://www.youtube.com/watch?v=AgkfIQ4IGaM.
1 Optimization For Any Arbitary Neuron
2 Three Forms of Regularizations
This paper proposed some more explicit forms of regularization. Different hyperparameter combinations of these regularizations produce different styles of reconizable images.
.2.1 Clipping Pixels with Small Contribution
Contribution can be measured as
Where Ã(d, i, j) is the activation with a certain pixel of X be set to zero.
However, this approach is slow since it needs to forward pass for every pixel. So we approximate Ã(d, i, j) by
So
Finally, we sum across channels to find pixels with small distribution in either direction, positive or negative, and set those pixels to zero by thresholding.
2.2 Blur X a Bit
Producing images via gradient ascent tends to produce examples with high frequency information. While these images cause high activations, they are neither realistic nor interpretable. As a result, we blur the image with Gaussian kernel every several steps
due to the computational effectiveness.
2.3 Take Any Pixel With Small Norm to Zero
CNN tends to give X non-zero everywhere to contribute in whatever small way they can to ultimately raise the chosen unit’s activation. We instead show only the main object, letting other regions be exactly zero by thresholding if they are not needed.
3 References
[1]. Deep Visualization Toolbox. https://www.youtube.com/watch?v=AgkfIQ4IGaM.
相关文章推荐
- [深度学习论文笔记][Image Classification] ImageNet Classification with Deep Convolutional Neural Networks
- [深度学习论文笔记][Human Pose Estimation] DeepPose: Human Pose Estimation via Deep Neural Networks
- Understanding Neural Networks Through Deep Visualization论文
- [深度学习论文笔记][Visualizing] Visualizing and Understanding Convolutional Networks
- [深度学习论文笔记][Visualizing] Deep Inside Convolutional Networks Visualising Image Classification
- [深度学习论文笔记][Recurrent Neural Networks] Visualizing and Understanding Recurrent Networks
- 【深度学习论文笔记:Recognition】:Deep Neural Networks for Object Detection
- [深度学习论文笔记][Weight Initialization] Understanding the difficulty of training deep feedforward neural
- [深度学习论文笔记][Neural Arts] Inceptionism: Going Deeper into Neural Networks
- [深度学习论文笔记][Video Classification] Large-scale Video Classification with Convolutional Neural Networks
- [深度学习论文笔记][Image Classification] Very Deep Convolutional Networks for Large-Scale Image Recognitio
- [深度学习论文笔记][Image Classification] Identity Mappings in Deep Residual Networks
- [深度学习论文笔记][Image Reconstruction] Understanding Deep Image Representations by Inverting Them
- [深度学习论文笔记][Video Classification] Beyond Short Snippets: Deep Networks for Video Classification
- Understanding Neural Networks Through Deep Visualization
- [深度学习论文笔记][arxiv 1711] Non-local Neural Networks
- Understanding Neural Networks Through Deep Visualization实现
- 深度学习入门笔记:Fast Image Search with Deep Convolutional Neural Networks and Efficient Hashing Codes
- [深度学习论文笔记][Adversarial Examples] Intriguing properties of neural networks
- 深度学习论文笔记:Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition