1. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
Deep Learning
Convolutional Neural Networks
Christian S. Perone
christian.perone@gmail.com
2. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
WHO AM I
Christian S. Perone
Software Designer
Blog
http://blog.christianperone.com
Open-source projects
https://github.com/perone
Twitter @tarantulae
3. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
AGENDA
DEEP LEARNING
Introduction
Traditional vs Deep learning
ImageNet Challenge
Deep learning in art
NEURAL NETWORKS
Neural network basics
Making it possible
CONVOLUTIONAL NEURAL NETWORKS
Architecture overview
Convolutional layer
Pooling layer
Dense layers and classification
Deep CNNs
Important ideas
Transfer learning
INTERESTING CASES
Recommendation
Natural language processing
Image/video processing
Q&A
4. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
Section I
DEEP LEARNING
5. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
WHAT IS DEEP LEARNING ?
Multiple definitions, however, these definitions have in common:
Multiple layers of processing units;
Supervised or unsupervised learning of feature representations
in each layer, with the layers forming a hierarchy from
low-level to high-level features.
6. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
COMPOSITIONAL DATA
NATURAL DATA
IS COMPOSITIONAL.
7. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
COMPOSITIONAL DATA
Image
Source: Convolutional Deep Belief Networks. Honglak Lee, et. al.
8. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
COMPOSITIONAL DATA
Sound
Source: Large Scale Deep Learning. Jeff Dean, joint work with Google.
9. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
ONE SLIDE INTRO TO MACHINE LEARNING
Source: Scikit-Learn (scikit-learn.org)
10. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
TRADITIONAL VS DEEP LEARNING
For many years, we developed feature extractors.
Source: Deep Learning Methods for Vision (Honglak Lee)
11. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
TRADITIONAL VS DEEP LEARNING
Feature extractors, required:
Expert knowledge
Time-consuming hand-tuning
In industrial applications, this is 90% of the time
Sometimes are problem specific
12. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
TRADITIONAL VS DEEP LEARNING
Feature extractors, required:
Expert knowledge
Time-consuming hand-tuning
In industrial applications, this is 90% of the time
Sometimes are problem specific
But, what if we could learn feature extractors ?
13. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
TRADITIONAL VS DEEP LEARNING
TRADITIONAL APPROACH
The traditional approach uses fixed feature extractors.
14. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
TRADITIONAL VS DEEP LEARNING
TRADITIONAL APPROACH
The traditional approach uses fixed feature extractors.
DEEP LEARNING APPROACH
Deep Learning approach uses trainable feature extractors.
15. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
TRADITIONAL VS DEEP LEARNING
Source: Lee et.al., ICML2009
16. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
IMAGENET
Source: t-SNE visualization of CNN codes. Andrej Karpathy
≈ 20.000 object classes
≈ 14 million images
18. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
IMAGE CLASSIFICATION
Source: We’ve Been Dressing Animals Up as People Way Before the Internet. Jes Greene.
Image classification, can get really hard.
19. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
IMAGENET CHALLENGE
Source: Musings on Deep Learning. Li Jiang.
20. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
DEEP DREAMS
Source: Google Inceptionism
21. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
ART STYLE
Source: A Neural Algorithm of Artistic Style. Leon A. Gatys et. al.
22. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
ART STYLE
Source: A Neural Algorithm of Artistic Style. Leon A. Gatys et. al.
23. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
ART STYLE
Source: A Neural Algorithm of Artistic Style. Leon A. Gatys et. al.
24. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
ART STYLE
Source: A Neural Algorithm of Artistic Style. Leon A. Gatys et. al.
25. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
Section II
NEURAL NETWORKS
26. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
NEURAL NETWORK ARCHITECTURE
Source: Neural Networks and Deep Learning. Michael Nielsen.
Source: Practical Deep N. Networks. Yuhuang Hu et. al.
27. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
MNIST DIGITS CLASSIFICATION
Segmented digits
MNIST digit format (28 x 28 = 784 pixels)
Source: Neural Networks and Deep Learning. Michael Nielsen.
28. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
NEURAL NETWORK ARCHITECTURE
Source: Neural Networks and Deep Learning. Michael Nielsen.
2.225 of 10.000 test images (22.25 % accuracy)
29. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
NEURAL NETWORK ARCHITECTURE
Source: Neural Networks and Deep Learning. Michael Nielsen.
2.225 of 10.000 test images (22.25 % accuracy)
An SVM classifier can get 9.435 of 10.000 ( % 94.35)
30. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
NEURAL NETWORK ARCHITECTURE
Source: Neural Networks and Deep Learning. Michael Nielsen.
2.225 of 10.000 test images (22.25 % accuracy)
An SVM classifier can get 9.435 of 10.000 ( % 94.35)
SVM with hyperparameter optimization can get 98.5%
31. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
NEURAL NETWORK ARCHITECTURE
Can we do better ?
32. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
NEURAL NETWORK ARCHITECTURE
Can we do better ?
In fact, yes. The current record is from 2013 and it classifies 9.979
of 10.000 images correctly. The performance is human-equivalent
(or better).
33. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
NEURAL NETWORK ARCHITECTURE
Can we do better ?
In fact, yes. The current record is from 2013 and it classifies 9.979
of 10.000 images correctly. The performance is human-equivalent
(or better).
Source: Neural Networks and Deep Learning. Michael Nielsen.
Neural networks can accurately classify all but 21 of the 10,000 test
images.
34. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
WHAT CHANGED ?
For approximately 20 years, attempts were made to train deeper
neural networks (with more than one hidden layer), however rarely
with benefits (vanishing gradient).
35. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
WHAT CHANGED ?
In 2006, a major breakthrough was made in deep architectures,
following three key principles:
Unsupervised learning of representations is used to pre-train
each layer
36. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
WHAT CHANGED ?
In 2006, a major breakthrough was made in deep architectures,
following three key principles:
Unsupervised learning of representations is used to pre-train
each layer
Unsupervised training of one layer at a time, on top of the
previously trained ones. The representation learned at each
level is the input for the next layer.
37. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
WHAT CHANGED ?
In 2006, a major breakthrough was made in deep architectures,
following three key principles:
Unsupervised learning of representations is used to pre-train
each layer
Unsupervised training of one layer at a time, on top of the
previously trained ones. The representation learned at each
level is the input for the next layer.
Use supervised training to fine-tune all the layers (in addition
to one or more additional layers that are dedicated to
producing predictions).
38. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
WHAT CHANGED ?
After the 2006 breakthrough, a lot of ideas were also developed.
Nowadays, pre-training is almost obsolete.
39. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
WHAT CHANGED ?
After the 2006 breakthrough, a lot of ideas were also developed.
Nowadays, pre-training is almost obsolete.
New activation functions
40. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
WHAT CHANGED ?
After the 2006 breakthrough, a lot of ideas were also developed.
Nowadays, pre-training is almost obsolete.
New activation functions
Regularization methods
41. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
WHAT CHANGED ?
After the 2006 breakthrough, a lot of ideas were also developed.
Nowadays, pre-training is almost obsolete.
New activation functions
Regularization methods
Initialization methods
42. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
WHAT CHANGED ?
After the 2006 breakthrough, a lot of ideas were also developed.
Nowadays, pre-training is almost obsolete.
New activation functions
Regularization methods
Initialization methods
Data augmentation
43. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
WHAT CHANGED ?
After the 2006 breakthrough, a lot of ideas were also developed.
Nowadays, pre-training is almost obsolete.
New activation functions
Regularization methods
Initialization methods
Data augmentation
Optimization techniques
44. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
WHAT CHANGED ?
Another reason on why Deep Learning is possible, is the availability
of lots of data (i.e. ImageNet).
45. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
WHAT CHANGED ?
Another reason on why Deep Learning is possible, is the availability
of lots of data (i.e. ImageNet).
GPGPU also plays an important role on this. For instance, an
NVIDIA GPU (1 Tesla K40 GPU) training a 7 layer Convolutional
Neural Network is nearly 9x faster than CPU.
Convolutions — 80-90% of execution time
Pooling
Activations
46. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
WHAT CHANGED ?
Companies are working on solutions for Deep Learning
acceleration:
47. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
WHAT CHANGED ?
Companies are working on solutions for Deep Learning
acceleration:
NVIDIA
NVIDIA created a entire plaftorm stack dedicated to work with Deep
Learning, called DIGITS. Their GPUs are widely used in Deep Learning.
48. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
WHAT CHANGED ?
Companies are working on solutions for Deep Learning
acceleration:
NVIDIA
NVIDIA created a entire plaftorm stack dedicated to work with Deep
Learning, called DIGITS. Their GPUs are widely used in Deep Learning.
AMAZON
Amazon AWS also create EC2 instances with NVIDIA GPUs (with 4GB
of memory and 1536 CUDA cores). Lots of AMIs with Deep Learning
software ecosystem already installed.
49. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
WHAT CHANGED ?
Companies are working on solutions for Deep Learning
acceleration:
NVIDIA
NVIDIA created a entire plaftorm stack dedicated to work with Deep
Learning, called DIGITS. Their GPUs are widely used in Deep Learning.
AMAZON
Amazon AWS also create EC2 instances with NVIDIA GPUs (with 4GB
of memory and 1536 CUDA cores). Lots of AMIs with Deep Learning
software ecosystem already installed.
MICROSOFT
Microsoft announced that it will offer NVIDIA GPUs on its Azure cloud
platform.
50. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
Section III
CONVOLUTIONAL NEURAL NETWORKS
51. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
CONVOLUTIONAL NEURAL NETWORKS
Convolutional Neural Networks (or convnets) are based on the
following principles:
Local receptive fields
Shared weights
Pooling (or down-sampling)
This special neural network architecture takes advantage of the
spatial structure of data.
52. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
CONVOLUTIONAL NEURAL NETWORKS
Convolutional Neural Networks (or convnets) are based on the
following principles:
Local receptive fields
Shared weights
Pooling (or down-sampling)
This special neural network architecture takes advantage of the
spatial structure of data.
Source: Deeply-Supervised Nets. Zhuowen Tu.
53. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
LOCAL CONNECTIVITY
Let’s take the MNIST digits images as input of our convnet. These
images are 28x28 pixels:
28x28 image Local connectivity (5x5)
Source: Neural Networks and Deep Learning. Michael Nielsen.
54. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
LOCAL CONNECTIVITY
55. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
LOCAL CONNECTIVITY
By “sliding” it, we create a feature map of or 24x24 neurons in the
hidden layer. We can also have a different stride and padding.
56. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
SHARED WEIGHTS
In this local receptive field, Convolutional Neural Networks use the
same shared weights for each of the 24x24 hidden neurons. This
means that we have a great advantage of parameter reduction, for
instance, for a 5x5 receptive field, we’ll need only 25 shared
weights1.
1
Excluding the bias
57. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
SHARED WEIGHTS
In this local receptive field, Convolutional Neural Networks use the
same shared weights for each of the 24x24 hidden neurons. This
means that we have a great advantage of parameter reduction, for
instance, for a 5x5 receptive field, we’ll need only 25 shared
weights1.
20 feature maps using 5x5 — 20*26 = 520 weights
A fully connected first layer, with 784=28*28 input neurons,
and a relatively modest 30 hidden neurons, would produce
784*30 = 23.520 weights, more than 40 times as many
parameters as the convolutional layer.
1
Excluding the bias
58. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
CONVOLUTIONAL LAYER
The shared weights and bias are called kernel or filter.
Convolutional layers provides
translation invariance. Since these
filters works on every part of the
image, they are “searching” for
the same feature everywhere in
the image.
60. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
POOLING LAYER
Pooling layers are usually present after a convolutional layer. They
provide a down-sampling of the convolution output.
In the example above, a 2x2 region is being used as input of the
pooling.
61. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
POOLING LAYER
There are different types of pooling, the most used is the
max-pooling and average pooling:
Pooling layers downsamples the volume spatially, reducing small
translations of the features. They also provide a parameter reduction.
Source: CS231n Convolutional Neural Networks for Visual Recognition.
62. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
POOLING LAYER
Max-pooling is how the network asks whether a feature is found
anywhere in some region of the image. After that, it will lose the
exact position.
64. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
CLASSIFICATION
As you can see, we then add a dense fully-connected layer (usually
using softmax) at the end of the neural network in order to get
predictions for the problem we’re working on (10 classes, 10 digits).
65. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
GOING DEEPER
We have defined all the components required to create a
Convolutional Neural Network, but you’ll rarely see a shallow
convnet like that.
66. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
GOING DEEPER
Actually, experiments demonstrated that the replication of
convolutional + pooling layers produces better results the deeper
you go. Winners of ImageNet challenge, have more than 15 layers
(VGGNet has 19 layers).
67. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
GOING DEEPER
Source: Rotation-invariant convolutional neural networks for galaxy morphology prediction. Sander Dieleman et. al.
Galaxy Zoo best performing network (winner of the challenge).
68. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
DROPOUT TECHNIQUE
Source: “Dropout: A Simple Way to Prevent Neural Networks from Overfitting”. Nitish Srivastava et. al.
The dropout technique helps with the overfitting, specially on dense
layers. Drop occur only at training time, not on test time.
69. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
ACTIVATION FUNCTIONS
Source: Big Data Analytics. Fei Wang.
ReLu helps with the vanishing gradient problem
ReLu generates sparsity
70. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
DATA AUGMENTATION
Data augmentation can help with overfitting and will certainly
improve improve results.
Source: Rotation-invariant convolutional neural networks for galaxy morphology prediction. Sander Dieleman et. al.
71. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
DATA AUGMENTATION
Data augmentation can help with overfitting and will certainly
improve improve results.
Source: Rotation-invariant convolutional neural networks for galaxy morphology prediction. Sander Dieleman et. al.
Small rotations
Small translation
Scaling
Flipping
Brightness
Noise
72. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
TRANSFER LEARNING
Features learned by Convolutional Neural Networks on large dataset
problem (i.e. ImageNet), can be helpful on different problems. It’s very
common to pre-train a convnet on ImageNet and then use it as a fixed
feature extractor or as initialization.
CONVNETS AS FEATURE EXTRACTORS
We can remove the last layer and then use these features to extract features,
these features are very useful features for classification. Some people use
these features with LSH (locality-sensitive hashing) to scale large databases
for image search. You can also use these features as input for a SVM
classifier for instance.
73. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
TRANSFER LEARNING
Features learned by Convolutional Neural Networks on large dataset
problem (i.e. ImageNet), can be helpful on different problems. It’s very
common to pre-train a convnet on ImageNet and then use it as a fixed
feature extractor or as initialization.
FINE-TUNING THE CONVNETS
You can use a pre-trained convnet to continue its training on your data
and thus fine-tune the weights for your problem. First layers of a convnet
contains generic features (i.e. edge detectors, etc.) that should be helpful in
many tasks. Deeper layers becomes progressively specific to the details of
the classes of the original problem.
74. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
Section IV
INTERESTING CASES
75. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
MUSIC RECOMMENDATION
Source: Recommending music on Spotify with deep learning. Sander Dieleman.
This is an example architecture from Spotify, using Convolutional
Neural Network for music recommendation.
76. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
MUSIC RECOMMENDATION
Learned filters at first convolutional layer.
The time axis is horizontal, the frequency axis is vertical (frequency
increases from top to bottom)
77. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
NATURAL LANGUAGE PROCESSING
Source: Text understanding from scratch. Xiang Zhang, Yann LeCun.
Deciding if a review posted on Amazon is positive or negative with
96% accuracy, and predict the actual number of stars with 73%
accuracy.
78. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
WORD2VEC
Source: Distributed Representations of Sentences and Documents. Quoc Le, Tomas Mikolov.
Word vectors (trained with up to hundreds of billions of words).
79. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
WORD2VEC
Source: Distributed Representations of Sentences and Documents. Quoc Le, Tomas Mikolov.
Word vectors (trained with up to hundreds of billions of words).
With nice properties:
v(’Paris’) - v(’France’) + v(’Italy’ ) ≈ v(’Rome’)
v(’king’) - v(’man’) + v(’woman’) ≈ v(’queen’)
No deep learning and no convnet, but a great distributed representation
example.
81. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
DOC2VEC
Sentiment analysis
Source: Distributed Representations of Sentences and Documents. Quoc Le, Tomas Mikolov.
82. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
INTERESTING FRAMES
Google recently put in production a Deep Neural Network to
improve YouTube video thumbnails.
Source: Google Research.
83. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
SPATIAL TRANSFORMER NETWOKRS
Spatial Transformer Networks can learn transformations.
Source: Spatial Transformer Networks. Max Jaderberg, et. al.
84. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
DEEPFACE BY FACEBOOK
Source: DeepFace: Closing the Gap to Human-Level Performance in Face Verification. Yaniv Taigman, et. al.
85. DEEP LEARNING NEURAL NETWORKS CONVOLUTIONAL NEURAL NETWORKS INTERESTING CASES Q&A
Section V
Q&A