A short presentation on Capsules Neural network

1. See discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/326866655
Capsule neural networks
Presentation · August 2018
CITATIONS
0
READS
8
1 author:
Some of the authors of this publication are also working on these related projects:
Wearable Clinic:(Adaptive Sensing & Signal Processing using wearable sensors ) View project
Gait Sensor Fusion View project
Tahmina Zebin
The University of Manchester
17 PUBLICATIONS 19 CITATIONS
SEE PROFILE
All content following this page was uploaded by Tahmina Zebin on 12 August 2018.
The user has requested enhancement of the downloaded file.

2. Capsule Neural Networks
TAHMINA.ZEBIN@MANCHESTER.AC.UK

3. Agenda
Part 1: Capsule Networks (CapsNet)
Part 2: Dynamic Routing Between Capsules
Part 3: CapsNet Architecture
Part 4: CapsNet Comparison with CNN : a demo android App

4. Part 1: Capsule Networks (CapsNet)
 The concept of capsules first introduced in 2011 by Geoffrey
Hinton, et al., in Transforming Autoencoders [1].
 In November 2017, Sara Sabour, Nicholas Frosst, and
Geoffrey Hinton in Dynamic Routing between Capsules [2],
where they introduced a CapsNet architecture that reached
state-of-the-art performance on MNIST dataset.
 Updated representation in Matrix capsules with EM routing
[3] (2018), 6th International Conference on Learning
Representations.

5. Challenges with convolutional networks..
 CNN is good at detecting features but less effective at exploring the spatial relationships among
features (perspective, size, orientation).
 CNN is also vulnerable to adversaries by simply move, rotate or resize individual features.
 It requires a large volume of training data to cover different variants and to avoid overfitting.
http://www.evolvingai.org/files/DNNsEasilyFooled_cvpr15.pdfSource :https://jhui.github.io/2017/11/03/Dynamic-Routing-Between-Capsules/

6. Capsules
A capsule is a group of neurons whose output represents different properties of the same
entity.
Capsules output a vector instead of a single scaler value.
In Matrix representation: VectorMatrix; Activity Vector Pose matrix + Activity Probability
A capsule captures the likelihood of a feature and its variant (orientation, size, perspective).
Uses the concept of Equivariance.
Image Source :https://jhui.github.io/2017/11/03/Dynamic-Routing-Between-Capsules/

7. Part 2: Dynamic Routing Between
Capsules
Dynamic routing groups capsules to form a parent capsule, and it calculates the capsule’s
output.
Source: Sara Sabour, Nicholas Frosst, Geoffrey Hinton

8. Routing by Agreement

9. Part 3: Capsnet Architecture
Source: Sara Sabour, Nicholas Frosst, Geoffrey Hinton

10. Part 4:
CapsNet
Comparison
with CNN :
a demo
android
App

11. Pros and Cons
Pros: Equivariance
Built in interpretability
Adversial Robustness
Cons: Reproducability
Computational performance
Routing process (with inner loops)

12. References & Further Reading
Founding papers by Hinton et al.
◦ Matrix capsules with EM routing - Hinton, G. E., Sabour, S. and Frosst, N. (2018), 6th International
Conference on Learning Representations.
https://openreview.net/pdf?id=HJWLfGWRb
◦ Dynamic Routing Between Capsules - Sabour, S., Frosst, N. and Hinton, G.E. (2017),
https://arxiv.org/abs/1710.09829
◦ Transforming Auto-encoders - Hinton, G. E., Krizhevsky, A. and Wang, S. D. (2011)
◦ https://github.com/sekwiatkowski/awesome-capsule-networks
◦ https://www.youtube.com/watch?v=YqazfBLLV4U&t=1259s
◦ https://jhui.github.io/2017/11/03/Dynamic-Routing-Between-Capsules/

13. Questions?
View publication statsView publication stats