Presentation made for the Developers World Forum 2018 Big Data and Artificial Intelligence Workshop.

1. Deep Learning in E-Commerce:
Applications and Challenges
A guide by Houda Bakir
houda.bakir@datavora.com

2. DL and ML Trends
2

4. Reebok Classic Leather Purple
$90.80
http://www.ebay.co.uk/itm/Reebok-Classic-Leather-
Womens-Suede-Purple-Red-Trainers-New-Shoes-All-
Sizes-/381019249763

6. AI, ML, DL
What a trend

7. AI, ML and DL

8. Evolution of ML

11. Classes of ML/DL

15. Product Matching

16. Pre Processing the Data
Use Text classification:
- Support Vector Machine (SVM)
- Convolution Neural Network (CNN)
In Order to FILTER the useful data.
Build time-series of targeted products.

17. Samsung Galaxy S8 64Gb

18. 2. Prediction H Bakir, G Chniti, H Zaher
E-Commerce price forecasting using LSTM neural networks
International Journal of Machine Learning and Computing 8 (2), 169-174

19. Use case: Shoes

20. Convolutional Neural Network

21. Clothes Classification
[Convolutional Neural Networks for Clothes Categories September
2015 doi: 10.1007/978-3-662-48570-5_12 ]

22. Use Case: Clothes Classification
Computer Vision For E-Commerce

23. Use Case: Clothes Classification

24. Use Case: Clothes Classification

25. A guide by Houda Bakir
houda.bakir@datavora.com
Convolutional Neural Networks

26. ➔ Introduction
➔ Convolution Neural Network
➔ The different Layers
➔ TensorFlow and Keras
26

27. Object Detection
27

28. - You will learn The Convolution Neural
Network
- How to implement it with Keras and
Tensorflow.
Agenda
28

29. Deep Neural Network
29
Fully Connected (FC) Neural Network

30. Inspiration of Convolution Networks
“Receptive fields, binocular interaction and functional architecture in the cat's visual cortex” 1968 - D. H. Hubel and T.
N. Wiesel
30

31. New Idea: Known unknowns => unknown unknowns
31
Neocognitron - Fukushima 1980

32. Convolutional Neural Network
Example of a CNN architecture:
LeNet-5 by LeCun in 1998
32

33. Convolutional neural networks (CNN)
Are similar to the neural networks. CNNs have weights, biases, and outputs through a
nonlinear activation.
33

34. A convolutional layer
34

35. 35

36. Convolutional Layer
stride=1
5 x 5 image
36
Filter 1
Filter 2
Repeat this for each filter

37. Convolution
37
1 0 0 0 0 1
0 1 0 0 1 0
0 0 1 1 0 0
1 0 0 0 1 0
0 1 0 0 1 0
0 0 1 0 1 0
6 x 6 image
1 -1 -1
-1 1 -1
-1 -1 1
Filter 1
-1 1 -1
-1 1 -1
-1 1 -1
Filter 2
…
…
These are the network
parameters to be learned.
Each filter detects a
small pattern (3 x 3).

38. Convolution
38
1 0 0 0 0 1
0 1 0 0 1 0
0 0 1 1 0 0
1 0 0 0 1 0
0 1 0 0 1 0
0 0 1 0 1 0
6 x 6 image
1 -1 -1
-1 1 -1
-1 -1 1
Filter 1
3 -1
stride = 1
Dot
product

39. Convolution
39
1 0 0 0 0 1
0 1 0 0 1 0
0 0 1 1 0 0
1 0 0 0 1 0
0 1 0 0 1 0
0 0 1 0 1 0
6 x 6 image
1 -1 -1
-1 1 -1
-1 -1 1
Filter 1
3 -3
stride = 2
Dot
product

40. Convolution
40
1 0 0 0 0 1
0 1 0 0 1 0
0 0 1 1 0 0
1 0 0 0 1 0
0 1 0 0 1 0
0 0 1 0 1 0
6 x 6 image
Filter 2
stride = 1
Dot
product
3 -1 -3 -1
-3 1 0 -3
-3 -3 0 1
3 -2 -2 -1
1 -1 -1
-1 1 -1
-1 -1 1

41. Convolution
41
1 0 0 0 0 1
0 1 0 0 1 0
0 0 1 1 0 0
1 0 0 0 1 0
0 1 0 0 1 0
0 0 1 0 1 0
6 x 6 image
stride = 1
Dot
product
-1 1 -1
-1 1 -1
-1 1 -1
3 -1 -3 -1
-3 1 0 -3
-3 -3 0 1
3 -2 -2 -1
-1 -1 -1 -1
-1 -1 -2 1
-1 -1 -2 1
-1 0 -4 3
Two 4 x 4 images
Forming 2 x 4 x 4 matrix
Feature
Map
Filter 2

42. Convolution v.s. Fully Connected
42
1 0 0 0 0 1
0 1 0 0 1 0
0 0 1 1 0 0
1 0 0 0 1 0
0 1 0 0 1 0
0 0 1 0 1 0
image
convolution
-1 1 -1
-1 1 -1
-1 1 -1
1 -1 -1
-1 1 -1
-1 -1 1
…
…
…
…
1 0 0 0 0 1
0 1 0 0 1 0
0 0 1 1 0 0
1 0 0 0 1 0
0 1 0 0 1 0
0 0 1 0 1 0
Fully-connected

43. 43
1 0 0 0 0 1
0 1 0 0 1 0
0 0 1 1 0 0
1 0 0 0 1 0
0 1 0 0 1 0
0 0 1 0 1 0
6 x 6 image
Filter 1
1
2
3
…
8
9
…
1
Only connect to
9 inputs, not
fully connected
4
:
10:
1
0
0
0
0
1
0
0
0
3
1 -1 -1
-1 1 -1
-1 -1 1
fewer parameters!

44. 44
1 0 0 0 0 1
0 1 0 0 1 0
0 0 1 1 0 0
1 0 0 0 1 0
0 1 0 0 1 0
0 0 1 0 1 0
6 x 6 image
Filter 1
…
1
0
0
1 -1 -1
-1 1 -1
-1 -1 1
1
:2
:3
:
…
7
:8
:9
:
…
1
4
:
10:
1
0
0
0
0
1
0
0
0
3
-1
Shared weights
fewer parameters!

45. Color image: RGB 3 channels
45
1 0 0 0 0 1
0 1 0 0 1 0
0 0 1 1 0 0
1 0 0 0 1 0
0 1 0 0 1 0
0 0 1 0 1 0
1 0 0 0 0 1
0 1 0 0 1 0
0 0 1 1 0 0
1 0 0 0 1 0
0 1 0 0 1 0
0 0 1 0 1 0
1 0 0 0 0 1
0 1 0 0 1 0
0 0 1 1 0 0
1 0 0 0 1 0
0 1 0 0 1 0
0 0 1 0 1 0
1 -1 -1
-1 1 -1
-1 -1 1 Filter 1
-1 1 -1
-1 1 -1
-1 1 -1
Filter 2
1 -1 -1
-1 1 -1
-1 -1 1
1 -1 -1
-1 1 -1
-1 -1 1
-1 1 -1
-1 1 -1
-1 1 -1
-1 1 -1
-1 1 -1
-1 1 -1
Color image

46. 46
The whole CNN
Fully Connected Feedforward network
cat dog …… Convolution
Max Pooling
Convolution
Max Pooling
Flattened
Can repeat
many times

47. Max Pooling
47

48. Max Pooling
48
3 -1 -3 -1
-3 1 0 -3
-3 -3 0 1
3 -2 -2 -1
-1 1 -1
-1 1 -1
-1 1 -1 Filter 2
-1 -1 -1 -1
-1 -1 -2 1
-1 -1 -2 1
-1 0 -4 3
1 -1 -1
-1 1 -1
-1 -1 1 Filter 1

49. 49
1 0 0 0 0 1
0 1 0 0 1 0
0 0 1 1 0 0
1 0 0 0 1 0
0 1 0 0 1 0
0 0 1 0 1 0
6 x 6 image
3 0
13
-1 1
30
2 x 2 image
Each filter
is a channel
Conv
Max
Pooling

50. 50
Convolution
Max Pooling
Convolution
Max Pooling
Can repeat
many times
A new image
The number of channels is the
number of filters
Smaller than the original
image
3 0
13
-1 1
30

51. 51
The whole CNN
Fully Connected Feedforward network
cat dog …… Convolution
Max Pooling
Convolution
Max Pooling
Flattened
Can repeat
many times

52. Flattening
52
3 0
13
-1 1
30 Flattened
3
0
1
3
-1
1
0
3
Fully Connected
Feedforward network

53. Visualizing Convolution Neural Network
53

54. Visualizing Neural Networks
54

55. Region-based Convolutional Networks (R-CNN)
55
Source: Mask R-CNN paper.

56. CNN improvement
56

57. Tensorflow
1. Powerful Deep Learning Framework
2. Keras
3. The approach is optional (Tensorflow Eager)
4. Tensorboard
57

58. CNN in Keras
58
Convolution
Max Pooling
Convolution
Max Pooling
Input
1 x 28 x 28
25 x 26 x 26
25 x 13 x 13
50 x 11 x 11
50 x 5 x 5
How many
parameters for each
filter?
How many
parameters
for each filter?
9
225=
25x9

59. CNN in Keras
59
Convolution
Max Pooling
Convolution
Input
1 x 28 x 28
25 x 26 x 26
25 x 13 x 13
50 x 11 x 11
Flattened
1250
Fully connected
feedforward network
Output
50 x 5 x 5
Max Pooling

60. CNN LOSS FUNCTIONS
Can be categorized in to the following categories:
1. Binary Classification(SVM hinge loss, Squared hinge
loss).
2. Identity Verification(Contrastive loss).
3. Multi-class Classification (Softmax loss, Expectation loss).
4. Regression (SSIM,`1 error, Euclidean loss)
60

61. 61

62. 62

63. Questions?
63