Inspirational applications of deep learning

1. Inspirational Applications
of Deep Learning
Sergey Shalnov, ADSKOM
sergey@adskom.com

2. 27.07.16 2

3. Learning visual representations [Andrea Vedaldi]
COMPUTER VISION
27.07.16 3

4. FEATURE EXTRACTION (CLASSIC APPROACH)
• Histogram of Oriented Gradients, HOG
27.07.16 4

5. CONVOLUTIONAL NETWORKS
• LeNet - Yann LeCun, 1998
• But need high computational resources
• It needs very high computational resources for 1998
27.07.16 5

6. Convolution with 3×3 Filter. Source
CONVOLUTION
27.07.16 6

7. CONVOLUTION
27.07.16 7
• AVERAGING EACH PIXEL WITH ITS NEIGHBORING VALUES BLURS AN IMAGE

8. CONVOLUTION
27.07.16 8
• AVERAGING EACH PIXEL WITH ITS NEIGHBORING VALUES BLURS AN IMAGE

9. CS231n Convolutional Neural Networks for Visual Recognition
POOLING
• Pooling layer downsamples the volume spatially, independently in each depth slice of the
input volume. Left: In this example, the input volume of size [224x224x64] is pooled with
filter size 2, stride 2 into output volume of size [112x112x64]. Notice that the volume
depth is preserved. Right: The most common downsampling operation is max, giving rise
to max pooling, here shown with a stride of 2. That is, each max is taken over 4 numbers
(little 2x2 square
27.07.16 9

10. CNN ARCHITECTURE
27.07.16 10

11. Zeiler and Fergus 2013, ”Visualizing and Understanding Convolutional Networks".
CNN FEATURES
• Learned CNN features.
• The features become more extended and complex deeper in the network.
27.07.16 11

12. Zeiler and Fergus 2013, ”Visualizing and Understanding Convolutional Networks".
CNN FEATURES
• Learned CNN features.
27.07.16 12

13. Zeiler and Fergus 2013, ”Visualizing and Understanding Convolutional Networks".
CNN FEATURES
•
27.07.16 13

14. GOOGLENET ARCHITECTURE (2014)
27.07.16 14

15. NATURAL IMAGE CLASSIFICATION - ImageNet
• Alex Krizhevsky , 2012
• 1.2M training images, 1000
classes
• Scored 15.3% Top-5 error
rate with 26.2% for the
second-best entry for
classification task
• CNNs trained with GPUs
• Demo
27.07.16 15

16. Learning and Transferring Mid-Level Image Representations using Convolutional Neural Networks
[Oquab et al. CVPR 2014]
TRANSFER LEARNING, IMAGE SEARCH
27.07.16 16

17. Deep Neural Networks for Object Detection [pdf], 2013
OBJECT DETECTION
27.07.16 17

18. Deep Visual-Semantic Alignments for Generating Image Descriptions [Andrej Karpathy, Li Fei-Fei]
IMAGE CAPTIONING
• Large CNN for the object
detection in the
photographs
• recurrent neural
network like an LSTM to
turn the labels into a
coherent sentence
• web demo
27.07.16 18

19. Image taken from Richard Zhang, Phillip Isola and Alexei A. Efros.
IMAGE COLORIZATION
• DEMO, VIDEO
27.07.16 19

20. RECURRENT NEURAL NETWORKS
• Sequence processing, memory
• A, looks at some input Xt and outputs a value Ht.
• A loop allows information to be passed from one step of the network to the next
27.07.16 20

21. RNN
• Problem - Long-Term
Dependencies
• “I grew up in France… I speak
fluent French.”
27.07.16 21

22. LSTM
• Adding the cell state
• Understanding LSTM
27.07.16 22

23. The Unreasonable Effectiveness of Recurrent Neural Networks [Andrej Karpathy]
TEXT GENERATION
27.07.16 23

24. TEXT GENERATION
27.07.16 24

25. • 3-layer RNN with 512 hidden nodes on each layer. Trained on the Shakespeare
27.07.16 25

26. Sampled (fake) algebraic geometry. Here's the actual pdf.
TEXT GENERATION
27.07.16 26

27. Generating Sequences With Recurrent Neural Networks [pdf], 2013
Demo
HANDWRITING GENERATION
27.07.16 27

28. t-SNE visualizations of word embeddings. Left: Number Region; Right: Jobs Region. From Turian et
al. (2010), see complete image.
WORD EMBEDDINGS
• A word embedding is a parameterized function mapping words in some language
to high-dimensional vectors (perhaps 200 to 500 dimensions).
• W(“cat")=(0.2, -0.4, 0.7, ...)
• W("table")=(0.0, 0.6, -0.1, ...)
27.07.16 28

29. What words have embeddings closest to a given word? From Collobertet al. (2011)
WORD EMBEDDINGS
27.07.16 29

30. From Mikolov et al.(2013a)
WORD EMBEDDINGS
• Relationship
• W(‘‘woman")−W(‘‘man") ≃ W(‘‘aunt")−W(‘‘uncle")
• W(‘‘woman")−W(‘‘man") ≃ W(‘‘queen")−W(‘‘king")
27.07.16 30

31. Relationship pairs in a word embedding. From Mikolov et al. (2013b).
WORD EMBEDDINGS
27.07.16 31

32. Introduction to Neural Machine Translation with GPUs
NEURAL MACHINE TRANSLATION
27.07.16 32

33. NEURAL MACHINE TRANSLATION (ENCODER)
• Step 1: A word to a one-hot vector.
27.07.16 33

34. NEURAL MACHINE TRANSLATION (ENCODER)
• Step 2: A one-hot vector to a continuous-space representation.
27.07.16 34

35. NEURAL MACHINE TRANSLATION (ENCODER)
• Step 3: Sequence summarization by a recurrent neural network.
27.07.16 35

36. [Sutskever et al., 2014].
NEURAL MACHINE TRANSLATION
• 2-D Visualization of Sentence Representations. Similar sentences are close
together in summary-vector space.
27.07.16 36

37. NEURAL MACHINE TRANSLATION (DECODER)
• Step 1: Computing the internal hidden state of the decoder.
27.07.16 37

38. NEURAL MACHINE TRANSLATION (DECODER)
• Step 2: Next word probability.
27.07.16 38

39. NEURAL MACHINE TRANSLATION (DECODER)
• Step 3: Sampling the next word. Кepeating until we select the end-of-sentence word (<eos>).
27.07.16 39

40. REINFORCEMENT LEARNING
27.07.16 40
• a set of environment states S;
• a set of actions A;
• rules of transitioning between states;
• rules that determine the scalar immediate reward of a transition;
• rules that describe what the agent observes.

41. Demo
DEEP REINFORCEMENT LEARNING
27.07.16 41

42. Deep Learning Flappy Bird, DEMO
FLAPPY BIRD
27.07.16 42

43. Questions?