Cs231n 2017 lecture12 Visualizing and Understanding

Fei-Fei Li & Justin Johnson & Serena Yeung Lecture 12 - May 16, 2017Fei-Fei Li & Justin Johnson & Serena Yeung Lecture 12 - May 16, 20171
Lecture 12:
Visualizing and Understanding

Fei-Fei Li & Justin Johnson & Serena Yeung Lecture 11 - May 10, 2017Fei-Fei Li & Justin Johnson & Serena Yeung Lecture 11 - May 10, 20172
Administrative
Milestones due tonight on Canvas, 11:59pm
Midterm grades released on Gradescope this week
A3 due next Friday, 5/26
HyperQuest deadline extended to Sunday 5/21, 11:59pm
Poster session is June 6

Fei-Fei Li & Justin Johnson & Serena Yeung Lecture 11 - May 10, 20173
Last Time: Lots of Computer Vision Tasks
Classification
+ Localization
Semantic
Segmentation
Object
Detection
Instance
Segmentation
CATGRASS, CAT,
TREE, SKY
DOG, DOG, CAT DOG, DOG, CAT
Single Object Multiple ObjectNo objects, just pixels This image is CC0 public domainThis image is CC0 public domain

This image is CC0 public domain
Class Scores:
1000 numbers
What’s going on inside ConvNets?
Input Image:
3 x 224 x 224
What are the intermediate features looking for?
Krizhevsky et al, “ImageNet Classification with Deep Convolutional Neural Networks”, NIPS 2012.
Figure reproduced with permission.

First Layer: Visualize Filters
AlexNet:
64 x 3 x 11 x 11
ResNet-18:
64 x 3 x 7 x 7
ResNet-101:
64 x 3 x 7 x 7
DenseNet-121:
64 x 3 x 7 x 7
Krizhevsky, “One weird trick for parallelizing convolutional neural networks”, arXiv 2014
He et al, “Deep Residual Learning for Image Recognition”, CVPR 2016
Huang et al, “Densely Connected Convolutional Networks”, CVPR 2017

Visualize the
filters/kernels
(raw weights)
We can visualize
filters at higher
layers, but not
that interesting
(these are taken
from ConvNetJS
CIFAR-10
demo)
layer 1 weights
layer 2 weights
layer 3 weights
16 x 3 x 7 x 7
20 x 16 x 7 x 7
20 x 20 x 7 x 7

FC7 layer
Last Layer
4096-dimensional feature vector for an image
(layer immediately before the classifier)
Run the network on many images, collect the
feature vectors

Last Layer: Nearest Neighbors
Test image L2 Nearest neighbors in feature space
4096-dim vector
Recall: Nearest neighbors
in pixel space
Figures reproduced with permission.

Last Layer: Dimensionality Reduction
Van der Maaten and Hinton, “Visualizing Data using t-SNE”, JMLR 2008
Figure copyright Laurens van der Maaten and Geoff Hinton, 2008. Reproduced with permission.
Visualize the “space” of FC7
feature vectors by reducing
dimensionality of vectors from
4096 to 2 dimensions
Simple algorithm: Principle
Component Analysis (PCA)
More complex: t-SNE

Last Layer: Dimensionality Reduction
Van der Maaten and Hinton, “Visualizing Data using t-SNE”, JMLR 2008
Figure reproduced with permission.
See high-resolution versions at
http://cs.stanford.edu/people/karpathy/cnnembed/

Visualizing Activations
Yosinski et al, “Understanding Neural Networks Through Deep Visualization”, ICML DL Workshop 2014.
Figure copyright Jason Yosinski, 2014. Reproduced with permission.
conv5 feature map is
128x13x13; visualize
as 128 13x13
grayscale images

Maximally Activating Patches
Pick a layer and a channel; e.g. conv5 is
128 x 13 x 13, pick channel 17/128
Run many images through the network,
record values of chosen channel
Visualize image patches that correspond
to maximal activations
Springenberg et al, “Striving for Simplicity: The All Convolutional Net”, ICLR Workshop 2015
Figure copyright Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, Martin Riedmiller, 2015;
reproduced with permission.

Occlusion Experiments
Mask part of the image before
feeding to CNN, draw heatmap of
probability at each mask location
Zeiler and Fergus, “Visualizing and Understanding Convolutional
Networks”, ECCV 2014
Boat image is CC0 public domain
Elephant image is CC0 public domain
Go-Karts image is CC0 public domain

Saliency Maps
Dog
How to tell which pixels matter for classification?
Simonyan, Vedaldi, and Zisserman, “Deep Inside Convolutional Networks: Visualising Image Classification Models
and Saliency Maps”, ICLR Workshop 2014.
Figures copyright Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman, 2014; reproduced with permission.

Saliency Maps
Dog
How to tell which pixels matter for classification?
Compute gradient of (unnormalized) class
score with respect to image pixels, take
absolute value and max over RGB channels

Saliency Maps

Saliency Maps: Segmentation without supervision
Rother et al, “Grabcut: Interactive foreground extraction using iterated graph cuts”, ACM TOG 2004
Use GrabCut on
saliency map

Intermediate Features via (guided) backprop
Zeiler and Fergus, “Visualizing and Understanding Convolutional Networks”, ECCV 2014
Pick a single intermediate neuron, e.g. one
value in 128 x 13 x 13 conv5 feature map
Compute gradient of neuron value with respect
to image pixels

Intermediate features via (guided) backprop
Pick a single intermediate neuron, e.g. one
value in 128 x 13 x 13 conv5 feature map
Compute gradient of neuron value with respect
to image pixels
Images come out nicer if you only
backprop positive gradients through
each ReLU (guided backprop)
ReLU
Figure copyright Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas
Brox, Martin Riedmiller, 2015; reproduced with permission.

Intermediate features via (guided) backprop
Figure copyright Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, Martin Riedmiller, 2015; reproduced with permission.

Visualizing CNN features: Gradient Ascent
(Guided) backprop:
Find the part of an
image that a neuron
responds to
Gradient ascent:
Generate a synthetic
image that maximally
activates a neuron
I* = arg maxI
f(I) + R(I)
Neuron value Natural image regularizer

score for class c (before Softmax)
zero image
1. Initialize image to zeros
Repeat:
2. Forward image to compute current scores
3. Backprop to get gradient of neuron value with respect to image pixels
4. Make a small update to the image

Simonyan, Vedaldi, and Zisserman, “Deep Inside Convolutional Networks: Visualising Image Classification
Models and Saliency Maps”, ICLR Workshop 2014.
Simple regularizer: Penalize L2
norm of generated image

Simonyan, Vedaldi, and Zisserman, “Deep Inside Convolutional Networks: Visualising Image Classification
Models and Saliency Maps”, ICLR Workshop 2014.

Figure copyright Jason Yosinski, Jeff Clune, Anh Nguyen, Thomas Fuchs, and Hod Lipson, 2014.
Reproduced with permission.

Better regularizer: Penalize L2 norm of
image; also during optimization
periodically
(1) Gaussian blur image
(2) Clip pixels with small values to 0
(3) Clip pixels with small gradients to 0

periodically
Figure copyright Jason Yosinski, Jeff Clune, Anh Nguyen, Thomas Fuchs, and Hod Lipson, 2014. Reproduced with permission.

periodically

Use the same approach to visualize intermediate features

Adding “multi-faceted” visualization gives even nicer results:
(Plus more careful regularization, center-bias)
Nguyen et al, “Multifaceted Feature Visualization: Uncovering the Different Types of Features Learned By Each Neuron in Deep Neural Networks”, ICML Visualization for Deep Learning Workshop 2016.
Figures copyright Anh Nguyen, Jason Yosinski, and Jeff Clune, 2016; reproduced with permission.

Nguyen et al, “Multifaceted Feature Visualization: Uncovering the Different Types of Features Learned By Each Neuron in Deep Neural Networks”, ICML Visualization for Deep Learning Workshop 2016.
Figures copyright Anh Nguyen, Jason Yosinski, and Jeff Clune, 2016; reproduced with permission.

Nguyen et al, “Synthesizing the preferred inputs for neurons in neural networks via deep generator networks,” NIPS 2016
Figure copyright Nguyen et al, 2016; reproduced with permission.
Optimize in FC6 latent space instead of pixel space:

Fooling Images / Adversarial Examples
(1) Start from an arbitrary image
(2) Pick an arbitrary class
(3) Modify the image to maximize the class
(4) Repeat until network is fooled

What is going on? Ian Goodfellow will explain

Fei-Fei Li & Justin Johnson & Serena Yeung Lecture 11 - May 10, 201737 37
DeepDream: Amplify existing features
Rather than synthesizing an image to maximize a specific neuron, instead
try to amplify the neuron activations at some layer in the network
Choose an image and a layer in a CNN; repeat:
1. Forward: compute activations at chosen layer
2. Set gradient of chosen layer equal to its activation
3. Backward: Compute gradient on image
4. Update image
Mordvintsev, Olah, and Tyka, “Inceptionism: Going Deeper into Neural
Networks”, Google Research Blog. Images are licensed under CC-BY
4.0

Fei-Fei Li & Justin Johnson & Serena Yeung Lecture 11 - May 10, 201738 38
Rather than synthesizing an image to maximize a specific neuron, instead
try to amplify the neuron activations at some layer in the network
Equivalent to:
I* = arg maxI
∑i
fi
(I)2
Mordvintsev, Olah, and Tyka, “Inceptionism: Going Deeper into Neural
Networks”, Google Research Blog. Images are licensed under CC-BY
4.0
Choose an image and a layer in a CNN; repeat:
1. Forward: compute activations at chosen layer
2. Set gradient of chosen layer equal to its activation
3. Backward: Compute gradient on image
4. Update image

Code is very simple but
it uses a couple tricks:
(Code is licensed under Apache 2.0)

Jitter image

Jitter image
L1 Normalize gradients

Jitter image
L1 Normalize gradients
Clip pixel values
Also uses multiscale processing for a fractal effect (not shown)

Sky image is licensed under CC-BY SA 3.0

Image is licensed under CC-BY 4.0

Feature Inversion
Given a CNN feature vector for an image, find a new image that:
- Matches the given feature vector
- “looks natural” (image prior regularization)
Mahendran and Vedaldi, “Understanding Deep Image Representations by Inverting Them”, CVPR 2015
Given feature vector
Features of new image
Total Variation regularizer
(encourages spatial smoothness)

Feature Inversion
Reconstructing from different layers of VGG-16
Mahendran and Vedaldi, “Understanding Deep Image Representations by Inverting Them”, CVPR 2015
Figure from Johnson, Alahi, and Fei-Fei, “Perceptual Losses for Real-Time Style Transfer and Super-Resolution”, ECCV 2016. Copyright Springer, 2016.
Reproduced for educational purposes.

Texture Synthesis
Given a sample patch of some texture, can we
generate a bigger image of the same texture?
Input
Output
Output image is licensed under the MIT license

Texture Synthesis: Nearest Neighbor
Generate pixels one at a time in
scanline order; form neighborhood
of already generated pixels and
copy nearest neighbor from input
Wei and Levoy, “Fast Texture Synthesis using Tree-structured Vector Quantization”, SIGGRAPH 2000
Efros and Leung, “Texture Synthesis by Non-parametric Sampling”, ICCV 1999

Texture Synthesis: Nearest Neighbor
Images licensed under the MIT license

Neural Texture Synthesis: Gram Matrix
Each layer of CNN gives C x H x W tensor of
features; H x W grid of C-dimensional vectors
This image is in the public domain.
w
H
C

Outer product of two C-dimensional vectors
gives C x C matrix measuring co-occurrence
w
H
C
C
C

Average over all HW pairs of vectors, giving
Gram matrix of shape C x C
w
H
C
C
C
Gram Matrix

Average over all HW pairs of vectors, giving
Gram matrix of shape C x C
w
H
C
C
C
Efficient to compute; reshape features from
C x H x W to =C x HW
then compute G = FFT

Gatys, Ecker, and Bethge, “Texture Synthesis Using Convolutional Neural Networks”, NIPS 2015
Figure copyright Leon Gatys, Alexander S. Ecker, and Matthias Bethge, 2015. Reproduced with permission.
Neural Texture Synthesis
1. Pretrain a CNN on ImageNet (VGG-19)
2. Run input texture forward through CNN,
record activations on every layer; layer i
gives feature map of shape Ci
× Hi
× Wi
3. At each layer compute the Gram matrix
giving outer product of features:
(shape Ci
× Ci
)
4. Initialize generated image from random
noise
5. Pass generated image through CNN,
compute Gram matrix on each layer
6. Compute loss: weighted sum of L2
distance between Gram matrices
7. Backprop to get gradient on image
8. Make gradient step on image
9. GOTO 5

× Hi
× Wi
(shape Ci
× Ci
)
noise
9. GOTO 5

Reconstructing texture from
higher layers recovers
larger features from the
input texture

Neural Texture Synthesis: Texture = Artwork
Texture synthesis
(Gram
reconstruction)
Figure from Johnson, Alahi, and Fei-Fei, “Perceptual
Losses for Real-Time Style Transfer and
Super-Resolution”, ECCV 2016. Copyright Springer, 2016.

Neural Style Transfer: Feature + Gram
Reconstruction
Feature
reconstruction
Texture synthesis
(Gram
reconstruction)
Figure from Johnson, Alahi, and Fei-Fei, “Perceptual
Losses for Real-Time Style Transfer and
Super-Resolution”, ECCV 2016. Copyright Springer, 2016.

Neural Style Transfer
Content Image Style Image
+
This image is licensed under CC-BY 3.0 Starry Night by Van Gogh is in the public domain

Content Image Style Image Style Transfer!
+ =
This image is licensed under CC-BY 3.0 Starry Night by Van Gogh is in the public domain This image copyright Justin Johnson, 2015. Reproduced with
permission.
Gatys, Ecker, and Bethge, “Image style transfer using convolutional neural networks”, CVPR 2016

Style
image
Content
image
Output
image
(Start with
noise)
Figure adapted from Johnson, Alahi, and Fei-Fei, “Perceptual Losses for Real-Time Style Transfer and
Super-Resolution”, ECCV 2016. Copyright Springer, 2016. Reproduced for educational purposes.

Style
image
Content
image
Output
image
Figure adapted from Johnson, Alahi, and Fei-Fei, “Perceptual Losses for Real-Time Style Transfer and
Super-Resolution”, ECCV 2016. Copyright Springer, 2016. Reproduced for educational purposes.

Figure copyright Justin Johnson, 2015.
Example outputs from
my implementation
(in Torch)

More weight to
content loss
More weight to
style loss

Larger style
image
Smaller style
image
Resizing style image before running style transfer
algorithm can transfer different types of features

Neural Style Transfer: Multiple Style Images
Mix style from multiple images by taking a weighted average of Gram matrices

Problem: Style transfer
requires many forward /
backward passes through
VGG; very slow!

Problem: Style transfer
requires many forward /
backward passes through
VGG; very slow!
Solution: Train another
neural network to perform
style transfer for us!

78
Fast Style Transfer (1) Train a feedforward network for each style
(2) Use pretrained CNN to compute same losses as before
(3) After training, stylize images using a single forward pass
Johnson, Alahi, and Fei-Fei, “Perceptual Losses for Real-Time Style Transfer and Super-Resolution”, ECCV 2016
Figure copyright Springer, 2016. Reproduced for educational purposes.

Fast Style Transfer
Slow SlowFast Fast
Johnson, Alahi, and Fei-Fei, “Perceptual Losses for Real-Time Style Transfer and Super-Resolution”, ECCV 2016
Figure copyright Springer, 2016. Reproduced for educational purposes.
https://github.com/jcjohnson/fast-neural-style

Fast Style Transfer
Ulyanov et al, “Texture Networks: Feed-forward Synthesis of Textures and Stylized Images”, ICML 2016
Ulyanov et al, “Instance Normalization: The Missing Ingredient for Fast Stylization”, arXiv 2016
Figures copyright Dmitry Ulyanov, Vadim Lebedev, Andrea Vedaldi, and Victor Lempitsky, 2016. Reproduced with
permission.
Concurrent work from Ulyanov et al, comparable results

Fast Style Transfer
Ulyanov et al, “Texture Networks: Feed-forward Synthesis of Textures and Stylized Images”, ICML 2016
Ulyanov et al, “Instance Normalization: The Missing Ingredient for Fast Stylization”, arXiv 2016
Figures copyright Dmitry Ulyanov, Vadim Lebedev, Andrea Vedaldi, and Victor Lempitsky, 2016. Reproduced with
permission.
Replacing batch normalization with Instance Normalization improves results

One Network, Many Styles
Dumoulin, Shlens, and Kudlur, “A Learned Representation for Artistic Style”, ICLR 2017.
Figure copyright Vincent Dumoulin, Jonathon Shlens, and Manjunath Kudlur, 2016; reproduced with permission.

One Network, Many Styles
Dumoulin, Shlens, and Kudlur, “A Learned Representation for Artistic Style”, ICLR 2017.
Figure copyright Vincent Dumoulin, Jonathon Shlens, and Manjunath Kudlur, 2016; reproduced with permission.
Use the same network for multiple
styles using conditional instance
normalization: learn separate scale
and shift parameters per style
Single network can blend styles after training

Summary
Many methods for understanding CNN representations
Activations: Nearest neighbors, Dimensionality reduction,
maximal patches, occlusion
Gradients: Saliency maps, class visualization, fooling
images, feature inversion
Fun: DeepDream, Style Transfer.

Next time: Unsupervised Learning
Autoencoders
Variational Autoencoders
Generative Adversarial Networks

Cs231n 2017 lecture12 Visualizing and Understanding

Recommended

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Cs231n 2017 lecture12 Visualizing and Understanding (20)

More from Yanbin Kong (6)

Recently uploaded (20)

Cs231n 2017 lecture12 Visualizing and Understanding