Slide 1
Slide 1 text
1
Demystifying the neural network
black box
Christopher Lennan (Senior Data Scientist)
@chris_lennan
Tanuj Jain (Data Scientist)
@tjainn #idealoTech
Slide 2
Slide 2 text
Motivation
2
Slide 3
Slide 3 text
3
Motivation
Convolutional Neural Networks (CNN)
• CNNs are at the heart of most computer vision applications these days
• They have also established themselves as successful for various NLP tasks
• But: It is notoriously difficult to interpret/understand predictions!
• Problematic, because
o Understanding is just as important as performance
o Sensitive applications (medical or autonomous driving) require a thorough understanding of
its inner workings
o CNNs can be improved if their predictions are properly understood
Slide 4
Slide 4 text
idealo Use Case
● Why are these misclassified as bathrooms?
4
Use Case: Identify the area of Hotel Property
Slide 5
Slide 5 text
Convolutional Neural Network (CNN)
5
source: Zhou et. al., “Learning Deep Features for Discriminative Localization”
Slide 6
Slide 6 text
Convolutional Neural Network (CNN)
6
source: http://cs231n.github.io/convolutional-networks/
Slide 7
Slide 7 text
Convolutional Neural Network (CNN)
7
source: Zhou et. al., “Learning Deep Features for Discriminative Localization”
Slide 8
Slide 8 text
Interpretability
8
Slide 9
Slide 9 text
9
Interpreting CNNs
• The available methods for interpreting CNNs can be broadly grouped into
1. Attribution methods
• attributing a classification result back to the input pixels
• output is usually a heatmap over input pixels that indicates how much each pixel
contributed to the classification outcome
2. Visualization methods
• Visualize the patterns that filters have learned
• Generate images that cause maximum neuron activation
• Visualize neuron interactions
Slide 10
Slide 10 text
Attribution methods
1. Perturbation based approaches - Occlude area of interest to test its effectiveness in changing
predictions
2. Gradient based approaches - Calculation of gradients of output w.r.t. some network variable
a. Saliency Maps
b. Guided Backpropagation
c. Deconvolution
3. Relevance Score approaches
a. Class Activation Map (CAM)
b. Grad- CAM
c. Layerwise Relevance Propagation (LRP)
Relative importance of areas of the input image for predictions
Slide 11
Slide 11 text
Visualization methods
Neurons and parameters can be visualized at all levels of the network:
11
Looking at the learned parameters of the neural network
source: https://distill.pub/2017/feature-visualization/
Slide 12
Slide 12 text
Coding Time!
12
Slide 13
Slide 13 text
Idealo Use Cases
13
Slide 14
Slide 14 text
CAM-Analysis
Slide 15
Slide 15 text
CAM-Analysis
Slide 16
Slide 16 text
Insights With CAM
Swimming Pool misclassified as Bathroom
16
CAM
Slide 17
Slide 17 text
Insights With CAM
Swimming Pool misclassified as Bathroom
17
CAM
Slide 18
Slide 18 text
Insights With CAM
Swimming Pool misclassified as Bathroom
18
CAM
Slide 19
Slide 19 text
Insights With CAM
Swimming Pool misclassified as Bathroom
Using rails to misidentify Pool as Bathroom.
19
Slide 20
Slide 20 text
Insights With CAM
Bathroom correct classification
20
CAM
Slide 21
Slide 21 text
Insights With CAM
Bathroom correct classification
21
CAM
Slide 22
Slide 22 text
Insights With CAM
Bathroom correct classification
22
CAM
Slide 23
Slide 23 text
Insights With CAM
Bathroom correct classification
Using faucets to correctly identify Bathroom.
23
Slide 24
Slide 24 text
Image Aesthetics
Slide 25
Slide 25 text
Convolutional filter visualisations
Layer 23 MobileNet original
MobileNet Aesthetic
25
Slide 26
Slide 26 text
Convolutional filter visualisations
Layer 51 MobileNet original
MobileNet Aesthetic
26
Slide 27
Slide 27 text
Convolutional filter visualisations
Layer 79 MobileNet original
MobileNet Aesthetic
27
Slide 28
Slide 28 text
Image-ATM
Python package we developed for tagging images
28
Slide 29
Slide 29 text
Check us out! #idealoTech
29
https://github.com/idealo https://medium.com/idealo-tech-blog
Slide 30
Slide 30 text
Data Science Team
30
Slide 31
Slide 31 text
Tanuj Jain
[email protected]
@tjainn
Christopher Lennan
[email protected]
@chris_lennan
31
Slide 32
Slide 32 text
THE END
32