Sorting waste in the right way is an important part of our circular economy. All too often people doubt in which bin they have to put their waste. What if an app helps you with sorting? By combining AI and IoT technology in a gamification app that adds fun for users, we were able to realize an advertising platform for companies.
Beyond Boundaries: Leveraging No-Code Solutions for Industry Innovation
Integrating AI techniques to build meaningful product concepts (by Jochem Grietens)
1. 1.1 Integrating AI techniques to build meaningful product concepts 1
CONFIDENTIAL Template Innovation Day 2018CONFIDENTIAL
INTEGRATING AI TECHNIQUES TO BUILD
MEANINGFUL PRODUCT CONCEPTS
Jochem Grietens
System engineer – A.I. passionate
jochem.grietens@verhaert.com
TRACK 1 - CIRCULAR ECONOMY FOR A BETTER WORLD
2. 1.1 Integrating AI techniques to build meaningful product concepts 2
CONFIDENTIAL
Introduction
3. 1.1 Integrating AI techniques to build meaningful product concepts 3
CONFIDENTIAL
1. Introduction
2. A recycling use case: Hackathon
3. Machine learning: What does it mean to “see” ?
4. How does a machine “see” ?
Content
4. 1.1 Integrating AI techniques to build meaningful product concepts 4
CONFIDENTIAL
“Fail fast, fail cheap, learn faster”
Understanding comes from running through the cycle.
Rapid prototyping credo
5. 1.1 Integrating AI techniques to build meaningful product concepts 5
CONFIDENTIAL
A recycling use case
Load Hackathon Case
6. 1.1 Integrating AI techniques to build meaningful product concepts 6
CONFIDENTIAL
aaa
Load hackathon 2018
• Theme A: Help The Community
• Theme B: Change behavior
• aim for a sustainable world
• live a healthier life
• become a better citizen
7. 1.1 Integrating AI techniques to build meaningful product concepts 7
CONFIDENTIAL
8. 1.1 Integrating AI techniques to build meaningful product concepts 8
CONFIDENTIAL
1. Business understanding
Hackathon process
9. 1.1 Integrating AI techniques to build meaningful product concepts 9
CONFIDENTIAL
Closing the loop in a circular economy
Challenges for companies:
• Little incentive to close the loop
• Companies have no control over the downstream process.
Business understanding
10. 1.1 Integrating AI techniques to build meaningful product concepts 10
CONFIDENTIAL
Sorting is hard
Sorting is boring
Sorting is a community effort
yellow for plastic and metal blue for paper green for glass
Business understanding
11. 1.1 Integrating AI techniques to build meaningful product concepts 11
CONFIDENTIAL
Concept generation
12. 1.1 Integrating AI techniques to build meaningful product concepts 12
CONFIDENTIAL
3. Data understanding
4. Data preparation
5. Modeling/app building
Hackathon process
13. 1.1 Integrating AI techniques to build meaningful product concepts 13
CONFIDENTIAL
14. 1.1 Integrating AI techniques to build meaningful product concepts 14
CONFIDENTIAL
Goals
Help people sort
Fun and engaging
Create a community
Help close the circular economy loop.
15. 1.1 Integrating AI techniques to build meaningful product concepts 15
CONFIDENTIAL
16. 1.1 Integrating AI techniques to build meaningful product concepts 16
CONFIDENTIAL
17. 1.1 Integrating AI techniques to build meaningful product concepts 17
CONFIDENTIAL
18. 1.1 Integrating AI techniques to build meaningful product concepts 18
CONFIDENTIAL
19. 1.1 Integrating AI techniques to build meaningful product concepts 19
CONFIDENTIAL
Hackathon process
6 Evaluation
7 Deployment
20. 1.1 Integrating AI techniques to build meaningful product concepts 20
CONFIDENTIAL
Architecture
21. 1.1 Integrating AI techniques to build meaningful product concepts 21
CONFIDENTIAL
Machine learning model
What does it mean to “see”
22. 1.1 Integrating AI techniques to build meaningful product concepts 22
CONFIDENTIAL
What does it mean to “see”
Primate visual system
23. 1.1 Integrating AI techniques to build meaningful product concepts 23
CONFIDENTIAL
What does it mean to “see”
24. 1.1 Integrating AI techniques to build meaningful product concepts 24
CONFIDENTIAL
“Seeing”:
• Recognizing (=classification)
• Localization
• Segmentation
• Understanding
What does it mean to “see”
25. 1.1 Integrating AI techniques to build meaningful product concepts 25
CONFIDENTIAL
Neural network to fit our goal
26. 1.1 Integrating AI techniques to build meaningful product concepts 26
CONFIDENTIAL
How does a machine “see” ?
27. 1.1 Integrating AI techniques to build meaningful product concepts 27
CONFIDENTIAL
Neural network
28. 1.1 Integrating AI techniques to build meaningful product concepts 28
CONFIDENTIAL
Neural network
yellow for plastic and metal
blue for paper
green for glass
29. 1.1 Integrating AI techniques to build meaningful product concepts 29
CONFIDENTIAL
1. Input Encoding
Neural network
30. 1.1 Integrating AI techniques to build meaningful product concepts 30
CONFIDENTIAL
Input Encoding
Picture = 3 arrays of numbers
Neural network
R
G
B
31. 1.1 Integrating AI techniques to build meaningful product concepts 31
CONFIDENTIAL
Output encoding
1. metal ( 1 0 0 )
2. paper ( 0 1 0 )
3. Glass ( 0 0 1 )
Neural network
33. 1.1 Integrating AI techniques to build meaningful product concepts 33
CONFIDENTIAL
The neuron
Neural network
34. 1.1 Integrating AI techniques to build meaningful product concepts 34
CONFIDENTIAL
Neural network
35. 1.1 Integrating AI techniques to build meaningful product concepts 35
CONFIDENTIAL
Neural network
36. 1.1 Integrating AI techniques to build meaningful product concepts 36
CONFIDENTIAL
Neural network
37. 1.1 Integrating AI techniques to build meaningful product concepts 37
CONFIDENTIAL
Convolutional neural network
Conv. neural networkInput
Output
0,9
0,1
0
yellow for plastic and metal blue for paper green for glass
38. 1.1 Integrating AI techniques to build meaningful product concepts 38
CONFIDENTIAL
Neural network training
Forward pass
39. 1.1 Integrating AI techniques to build meaningful product concepts 39
CONFIDENTIAL
Neural network training
0,9
0,1
0
Prediction
40. 1.1 Integrating AI techniques to build meaningful product concepts 40
CONFIDENTIAL
Neural network training
0,9
0,1
0
Prediction
1
0
0
Truth
COMPARE
41. 1.1 Integrating AI techniques to build meaningful product concepts 41
CONFIDENTIAL
Prediction Truth
Calculate the error
Neural network training
0,9
0,1
0
1
0
0
42. 1.1 Integrating AI techniques to build meaningful product concepts 42
CONFIDENTIAL
Neural network training
Backwards pass
Prediction Truth
0,9
0,1
0
1
0
0
43. 1.1 Integrating AI techniques to build meaningful product concepts 43
CONFIDENTIAL
Many layers stacked
Deep neural network
44. 1.1 Integrating AI techniques to build meaningful product concepts 44
CONFIDENTIAL
Specific WIRING of the first regions for image recognition
Deep convolutional neural network
45. 1.1 Integrating AI techniques to build meaningful product concepts 45
CONFIDENTIAL
• Convolution
• Fully connected
Convolutional neural network
46. 1.1 Integrating AI techniques to build meaningful product concepts 46
CONFIDENTIAL
Convolution = comparing to patterns
Convolution
47. 1.1 Integrating AI techniques to build meaningful product concepts 47
CONFIDENTIAL
=6600
Convolution
Feature learning
Great match
48. 1.1 Integrating AI techniques to build meaningful product concepts 48
CONFIDENTIAL
= 0
Convolution
Convolution
Bad match
49. 1.1 Integrating AI techniques to build meaningful product concepts 49
CONFIDENTIAL
Feature maps
50. 1.1 Integrating AI techniques to build meaningful product concepts 50
CONFIDENTIAL
Neural network LEARNS the features from examples
Convolutional neural network
51. 1.1 Integrating AI techniques to build meaningful product concepts 51
CONFIDENTIAL
Neural network LEARNS the features from examples
Convolutional neural network
52. 1.1 Integrating AI techniques to build meaningful product concepts 52
CONFIDENTIAL
Convolutional network
53. 1.1 Integrating AI techniques to build meaningful product concepts 53
CONFIDENTIAL
Flattening
54. 1.1 Integrating AI techniques to build meaningful product concepts 54
CONFIDENTIAL
Flattening
55. 1.1 Integrating AI techniques to build meaningful product concepts 55
CONFIDENTIAL
Classification
56. 1.1 Integrating AI techniques to build meaningful product concepts 56
CONFIDENTIAL
Classification
0
1
Tesla
Volkswagen
One Hot encoding:
1
0
0
1
Tesla
0,2
0,8
57. 1.1 Integrating AI techniques to build meaningful product concepts 57
CONFIDENTIAL
Convolutional neural network
58. 1.1 Integrating AI techniques to build meaningful product concepts 58
CONFIDENTIAL
TrashNet dataset : to train on 2527 labeled images
Way too little to learn a deep neural net “to see” AND “to recognize”
Training data
59. 1.1 Integrating AI techniques to build meaningful product concepts 59
CONFIDENTIAL
Convolutional neural network
Transfer learning = Learning to see on a different dataset.
60. 1.1 Integrating AI techniques to build meaningful product concepts 60
CONFIDENTIAL
Can we learn how to see on Cifar dataset:
60,000 32x32 color images in 10 different classes.
Learn how to recognize on the trashnet database
Transfer learning
61. 1.1 Integrating AI techniques to build meaningful product concepts 61
CONFIDENTIAL
Transfer learning
Freeze Finetune
62. 1.1 Integrating AI techniques to build meaningful product concepts 62
CONFIDENTIAL
Learning by heart vs. learning general features
Algorithm should have learned general features :
• Visual effects of metallic light reflection
• Can shape features
• Size vs environment
• …
Instead it learned:
Very specific features, like the shadow on the hand …
Overtraining
MODEL DOES NOT GENERALIZE WELL
63. 1.1 Integrating AI techniques to build meaningful product concepts 63
CONFIDENTIAL
Central ML problem
Performance on new data ?
Test it!
How to detect overfitting
Training and tuning the
models.
Check performance
during development.
Don’t touch until the very
end.
64. 1.1 Integrating AI techniques to build meaningful product concepts 64
CONFIDENTIAL
Detecting overfitting
“Performance training set >> test set likely overfitting.”
Dealing with overfitting
65. 1.1 Integrating AI techniques to build meaningful product concepts 65
CONFIDENTIAL
More data - Data augmentation
Dealing with overtraining
66. 1.1 Integrating AI techniques to build meaningful product concepts 66
CONFIDENTIAL
Dealing with overtraining
X1 X2
67. 1.1 Integrating AI techniques to build meaningful product concepts 67
CONFIDENTIAL
Achieved
• 70% accuracy on test set.
• 65% in real world validation.
Results
68. 1.1 Integrating AI techniques to build meaningful product concepts 68
CONFIDENTIAL
One group, five brands
Our services are marketed through 5 brands each
addressing specific missions in product development.
INTEGRATED PRODUCT DEVELOPMENT
ON-SITE
PRODUCT
DEVELOPMENT
DIGITAL
PRODUCTS
DEVELOPMENT
OPTICAL
PRODUCTS
DEVELOPMENT
VENTURING