SlideShare ist ein Scribd-Unternehmen logo
1 von 22
Background
Stack of Layers with Convolution, Subsampling and Nonlinearity Operations
Modern CNNs
 Convolution Layer
–Filtering of unimportant information and extraction of salient local feature.
 Subsampling Layer
–Introduction of local transition invariance, reduction of computation and enlargement of receptive field.
 Nonlinearity
–Increase of capacity to model more complex input-output relationship.
Mistake and Disaster
What Geoff. Hinton Says
“The pooling operation used in convolutional neural networks is a
big mistake and the fact that it works so well is a disaster.”
“Internal data representation of a convolutional neural network
does not take into account important spatial hierarchies
between simple and complex objects.”
 What our brain does when we see the world is ‘inverse graphics’
–Extraction of hierarchical representation of the visual information
–The representation is ‘equivariance’, not ‘invariance’ to changes in viewpoint
Computer Graphics vs Computer Vision
Inverse Graphical Approach
 Sub-sampling tries to make the neural activities invariant for small changes in viewpoint.
–This is a silly goal, motivated by the fact that the final label needs to be viewpoint-invariant.
 Its better to aim for equivariance: Changes in viewpoint lead to corresponding changes in neural activities.
–In the perceptual system, its the weights that code viewpoint-invariant knowledge, not the neural
activities.
Better Representation via Better Objective
Invariance vs Equivariance
Capsules
 Transitional invariance
–The output is unchanged by small shift of object or parts
 Reduction of input
–Allowing more types of feature in the next layer
 Bad fit to psychology of shape perception
–We have intrinsic coordinate frames to objects and use it while recognizing things.
 Solution to wrong problem
–We need equivariance not invariance. Disentangling rather than discarding.
 Fails to use underlying linear structure
–We need to make used of the natural linear manifold which will help us to extrapolate.
 Poor wary to do dynamic routing
–We need to route each part of the input to the neurons that know how to deal with it. Routing via most
active neuron is primitive.
Pros and Cons
Arguments on Pooling
 If a low-level part moves to a very different position it will be represented by a different capsule
–‘Place-coded’ equivariance
 If a part only moves a small distance it will be represented by the same capsule but the pose of outputs of
the capsule will change
–‘Rate-coded’ equivariance
 Higher-level capsules have bigger domains so low-level ‘place-coded’ equivariance gets converted into
high-level ‘rate-coded’ equivariance.
 Without sub-sampling, CNNs can give ‘place-coded’ equivariance for discrete translations.
Place-coding vs Rate-coding
Two types of equivariance
Representation
Image
Translated
Representation
Translated
Image
 Neuron
–The activation of a specific neuron
= the likelihood of specific feature in the input
image.
–We need specialized mechanism(e.g.
generative models) to make each neuron
represent parameters of feature.
Neuron vs Capsule
 Capsule
– A group of neurons which not only capture the
likelihood of specific features but also the
parameters related to the features.
–The magnitude of activity vector represents the
probability of detecting specific features and its
orientation represents its parameters or properties
(position, orientation, scale, deformation, color,
hue, texture …)
 Neuron
–The inputs of lower layers and the outputs of
upper layers are scalars.
–After aggregating inputs multiplied by weights,
nonlinearity function such as ReLU is applied.
Neuron vs Capsule
 Capsule
– The inputs of the lower layers and the outputs of
upper layers are vectors.
– Instead of using nonlinearity function such as
ReLU, we used squashing function.
 Prediction vector
–With the previous capsule output and transformation matrix , we computer prediction vector as
 The capsule output of next layer
–Then the capsule output of the next layer is computed as
, where
 Coupling coefficient
–Here, are called coupling coefficient which is trained with dynamic routing algorithm.
–We impose restriction that = 1 which is achieved by softmax function of relevancy or similarity
score which is initialized with zeros and progressively updated as follows(which reminds me of
Hebbian learning rule):
Participants in the Algorithm
Dynamic Routing Algorithm
 Routing Algorithm
–The pseudo-code of routing algorithm is as follows:
 Loss Function
–For each digit capsule , the loss function is margin loss as
where when digit is present and . The default
Routing Algorithm and Loss Function
Dynamic Routing Algorithm
 Model Architecture
 Reconstruction Regularization
–To encourage the digit capsule to encode instantiation parameters of the input digits, we add
reconstruction loss (with weight 0.0005) to margin loss using a decoder with 3 fully connected layers.
CapsNet for MNIST
The Architecture of Capsule Network
 Classification + Reconstruction in One Diagram
CapsNet for MNIST
The Architecture of Capsule Network
Experiments
 Prediction and Reconstruction Example
– = label, prediction, reconstruction target
 Classification Accuracy
MNIST Dataset
Hand-written Digit Classification
 Improvement of recognition performance on novel viewpoints
smallNORB Dataset
Object Recognition from Various Angle
Conclusion
 We can train better models with less training data which is quite beneficial for medical image analysis
where the annotations are scarce and expensive to get.
 We need to study more on how our brain works. Pure data-driven approach with overcomplex models
might be wasting some resource(both data and computations)
 The CapsNet is the combination of credit assignment mechanism of CNN and instantization parameter
learning of generative models.
 CapsNet is slower than CNN counterpart according to the routing update iterations.
 CpasNet has not yet proven its effectiveness in large-scale visual recognition problems such as
‘ImageNet’
 Each element in the activity vector does not always represent meaningful properties of input image
 Cosine similarity may not be the best measure of similarity between the predicted capsule output and
actual capsule output.
Departure from ‘Conventional Neural Networks’
Discussion
 Paper
– Original NIPS 2017 Paper : https://arxiv.org/abs/1710.09829
– ICLR 2018 Paper : https://openreview.net/pdf?id=HJWLfGWRb
 Video
– Geofrrey Hinton's talk - "What is wrong with convolutional neural nets?"
• https://www.youtube.com/watch?v=rTawFwUvnLE&feature=youtu.be
– Siraj Raval lecture - "Capsule Networks : An Improvement to Convolutional Networks"
• https://www.youtube.com/watch?v=VKoLGnq15RM&feature=youtu.be&t=13m59s
 Related News and Blog Posts
– https://medium.com/@pechyonkin/understanding-hintons-capsule-networks-part-i-intuition-b4b559d1159b
– https://hackernoon.com/what-is-a-capsnet-or-capsule-network-2bfbe48769cc
– https://www.nextobe.com/single-post/2017/11/02/CapsNetCapsule-Network
– https://www.wired.com/story/googles-ai-wizard-unveils-a-new-twist-on-neural-networks/
– https://jhui.github.io/2017/11/03/Dynamic-Routing-Between-Capsules/
 Implementation
– https://github.com/nishnik/CapsNet-PyTorch
– https://github.com/debarko/CapsNet-Tensorflow
– https://github.com/XifengGuo/CapsNet-Keras
– https://github.com/jaesik817/adv_attack_capsnet
References
hello@vuno.co
Putting the world’s medical data to work

Weitere ähnliche Inhalte

Was ist angesagt?

Machine Learning 3 - Decision Tree Learning
Machine Learning 3 - Decision Tree LearningMachine Learning 3 - Decision Tree Learning
Machine Learning 3 - Decision Tree Learningbutest
 
Machine Learning - Convolutional Neural Network
Machine Learning - Convolutional Neural NetworkMachine Learning - Convolutional Neural Network
Machine Learning - Convolutional Neural NetworkRichard Kuo
 
"Getting More from Your Datasets: Data Augmentation, Annotation and Generativ...
"Getting More from Your Datasets: Data Augmentation, Annotation and Generativ..."Getting More from Your Datasets: Data Augmentation, Annotation and Generativ...
"Getting More from Your Datasets: Data Augmentation, Annotation and Generativ...Edge AI and Vision Alliance
 
Convolutional Neural Networks
Convolutional Neural NetworksConvolutional Neural Networks
Convolutional Neural NetworksTianxiang Xiong
 
Brief intro : Invariance and Equivariance
Brief intro : Invariance and EquivarianceBrief intro : Invariance and Equivariance
Brief intro : Invariance and Equivariance홍배 김
 
Introduction to Capsule Networks
Introduction to Capsule NetworksIntroduction to Capsule Networks
Introduction to Capsule NetworksChia-Ching Lin
 
Brain Tumor Segmentation using Enhanced U-Net Model with Empirical Analysis
Brain Tumor Segmentation using Enhanced U-Net Model with Empirical AnalysisBrain Tumor Segmentation using Enhanced U-Net Model with Empirical Analysis
Brain Tumor Segmentation using Enhanced U-Net Model with Empirical AnalysisMD Abdullah Al Nasim
 
Introduction to Deep Learning
Introduction to Deep LearningIntroduction to Deep Learning
Introduction to Deep LearningOleg Mygryn
 
K-means Clustering
K-means ClusteringK-means Clustering
K-means ClusteringAnna Fensel
 
K MEANS CLUSTERING
K MEANS CLUSTERINGK MEANS CLUSTERING
K MEANS CLUSTERINGsingh7599
 
UNet-VGG16 with transfer learning for MRI-based brain tumor segmentation
UNet-VGG16 with transfer learning for MRI-based brain tumor segmentationUNet-VGG16 with transfer learning for MRI-based brain tumor segmentation
UNet-VGG16 with transfer learning for MRI-based brain tumor segmentationTELKOMNIKA JOURNAL
 
Ch 7 Knowledge Representation.pdf
Ch 7 Knowledge Representation.pdfCh 7 Knowledge Representation.pdf
Ch 7 Knowledge Representation.pdfKrishnaMadala1
 
Lecture_16_Self-supervised_Learning.pptx
Lecture_16_Self-supervised_Learning.pptxLecture_16_Self-supervised_Learning.pptx
Lecture_16_Self-supervised_Learning.pptxKarimdabbabi
 
Deep learning for medical imaging
Deep learning for medical imagingDeep learning for medical imaging
Deep learning for medical imaginggeetachauhan
 
Deformable ConvNets V2, DCNV2
Deformable ConvNets V2, DCNV2Deformable ConvNets V2, DCNV2
Deformable ConvNets V2, DCNV2HaiyanWang16
 

Was ist angesagt? (20)

Machine Learning 3 - Decision Tree Learning
Machine Learning 3 - Decision Tree LearningMachine Learning 3 - Decision Tree Learning
Machine Learning 3 - Decision Tree Learning
 
Machine Learning - Convolutional Neural Network
Machine Learning - Convolutional Neural NetworkMachine Learning - Convolutional Neural Network
Machine Learning - Convolutional Neural Network
 
"Getting More from Your Datasets: Data Augmentation, Annotation and Generativ...
"Getting More from Your Datasets: Data Augmentation, Annotation and Generativ..."Getting More from Your Datasets: Data Augmentation, Annotation and Generativ...
"Getting More from Your Datasets: Data Augmentation, Annotation and Generativ...
 
Convolutional Neural Networks
Convolutional Neural NetworksConvolutional Neural Networks
Convolutional Neural Networks
 
Lecture11 - neural networks
Lecture11 - neural networksLecture11 - neural networks
Lecture11 - neural networks
 
Kmeans
KmeansKmeans
Kmeans
 
Brief intro : Invariance and Equivariance
Brief intro : Invariance and EquivarianceBrief intro : Invariance and Equivariance
Brief intro : Invariance and Equivariance
 
Introduction to Capsule Networks
Introduction to Capsule NetworksIntroduction to Capsule Networks
Introduction to Capsule Networks
 
Brain Tumor Segmentation using Enhanced U-Net Model with Empirical Analysis
Brain Tumor Segmentation using Enhanced U-Net Model with Empirical AnalysisBrain Tumor Segmentation using Enhanced U-Net Model with Empirical Analysis
Brain Tumor Segmentation using Enhanced U-Net Model with Empirical Analysis
 
Introduction to Deep Learning
Introduction to Deep LearningIntroduction to Deep Learning
Introduction to Deep Learning
 
K-means Clustering
K-means ClusteringK-means Clustering
K-means Clustering
 
K MEANS CLUSTERING
K MEANS CLUSTERINGK MEANS CLUSTERING
K MEANS CLUSTERING
 
UNet-VGG16 with transfer learning for MRI-based brain tumor segmentation
UNet-VGG16 with transfer learning for MRI-based brain tumor segmentationUNet-VGG16 with transfer learning for MRI-based brain tumor segmentation
UNet-VGG16 with transfer learning for MRI-based brain tumor segmentation
 
Ch 7 Knowledge Representation.pdf
Ch 7 Knowledge Representation.pdfCh 7 Knowledge Representation.pdf
Ch 7 Knowledge Representation.pdf
 
Lecture_16_Self-supervised_Learning.pptx
Lecture_16_Self-supervised_Learning.pptxLecture_16_Self-supervised_Learning.pptx
Lecture_16_Self-supervised_Learning.pptx
 
image classification
image classificationimage classification
image classification
 
Deep learning for medical imaging
Deep learning for medical imagingDeep learning for medical imaging
Deep learning for medical imaging
 
Deformable ConvNets V2, DCNV2
Deformable ConvNets V2, DCNV2Deformable ConvNets V2, DCNV2
Deformable ConvNets V2, DCNV2
 
Clustering
ClusteringClustering
Clustering
 
Clustering
ClusteringClustering
Clustering
 

Ähnlich wie Dynamic Routing Between Capsules

CAPSULE NETWORK PERFORMANCE WITH AUTONOMOUS NAVIGATION
CAPSULE NETWORK PERFORMANCE WITH AUTONOMOUS NAVIGATIONCAPSULE NETWORK PERFORMANCE WITH AUTONOMOUS NAVIGATION
CAPSULE NETWORK PERFORMANCE WITH AUTONOMOUS NAVIGATIONijaia
 
CAPSULE NETWORK PERFORMANCE WITH AUTONOMOUS NAVIGATION
CAPSULE NETWORK PERFORMANCE WITH AUTONOMOUS NAVIGATIONCAPSULE NETWORK PERFORMANCE WITH AUTONOMOUS NAVIGATION
CAPSULE NETWORK PERFORMANCE WITH AUTONOMOUS NAVIGATIONgerogepatton
 
Capsule Network Performance with Autonomous Navigation
Capsule Network Performance with Autonomous Navigation Capsule Network Performance with Autonomous Navigation
Capsule Network Performance with Autonomous Navigation gerogepatton
 
Fcv bio cv_cottrell
Fcv bio cv_cottrellFcv bio cv_cottrell
Fcv bio cv_cottrellzukun
 
Fcv bio cv_cottrell
Fcv bio cv_cottrellFcv bio cv_cottrell
Fcv bio cv_cottrellzukun
 
Batch normalization presentation
Batch normalization presentationBatch normalization presentation
Batch normalization presentationOwin Will
 
Deep Neural Network DNN.docx
Deep Neural Network DNN.docxDeep Neural Network DNN.docx
Deep Neural Network DNN.docxjaffarbikat
 
Dynamic Routing Between Capsules.pdf
Dynamic Routing Between Capsules.pdfDynamic Routing Between Capsules.pdf
Dynamic Routing Between Capsules.pdfAbeerPareek1
 
Neural Architecture Search: Learning How to Learn
Neural Architecture Search: Learning How to LearnNeural Architecture Search: Learning How to Learn
Neural Architecture Search: Learning How to LearnKwanghee Choi
 
Deep learning from a novice perspective
Deep learning from a novice perspectiveDeep learning from a novice perspective
Deep learning from a novice perspectiveAnirban Santara
 
Machine learning Module-2, 6th Semester Elective
Machine learning Module-2, 6th Semester ElectiveMachine learning Module-2, 6th Semester Elective
Machine learning Module-2, 6th Semester ElectiveMayuraD1
 
Review-image-segmentation-by-deep-learning
Review-image-segmentation-by-deep-learningReview-image-segmentation-by-deep-learning
Review-image-segmentation-by-deep-learningTrong-An Bui
 
Dynamic routing between capsules - A brief presentation
Dynamic routing between capsules - A brief presentationDynamic routing between capsules - A brief presentation
Dynamic routing between capsules - A brief presentationRomain Sabathé
 
ML Module 3 Non Linear Learning.pptx
ML Module 3 Non Linear Learning.pptxML Module 3 Non Linear Learning.pptx
ML Module 3 Non Linear Learning.pptxDebabrataPain1
 
Neural Networks in Data Mining - “An Overview”
Neural Networks  in Data Mining -   “An Overview”Neural Networks  in Data Mining -   “An Overview”
Neural Networks in Data Mining - “An Overview”Dr.(Mrs).Gethsiyal Augasta
 

Ähnlich wie Dynamic Routing Between Capsules (20)

CAPSULE NETWORK PERFORMANCE WITH AUTONOMOUS NAVIGATION
CAPSULE NETWORK PERFORMANCE WITH AUTONOMOUS NAVIGATIONCAPSULE NETWORK PERFORMANCE WITH AUTONOMOUS NAVIGATION
CAPSULE NETWORK PERFORMANCE WITH AUTONOMOUS NAVIGATION
 
CAPSULE NETWORK PERFORMANCE WITH AUTONOMOUS NAVIGATION
CAPSULE NETWORK PERFORMANCE WITH AUTONOMOUS NAVIGATIONCAPSULE NETWORK PERFORMANCE WITH AUTONOMOUS NAVIGATION
CAPSULE NETWORK PERFORMANCE WITH AUTONOMOUS NAVIGATION
 
Capsule Network Performance with Autonomous Navigation
Capsule Network Performance with Autonomous Navigation Capsule Network Performance with Autonomous Navigation
Capsule Network Performance with Autonomous Navigation
 
Fcv bio cv_cottrell
Fcv bio cv_cottrellFcv bio cv_cottrell
Fcv bio cv_cottrell
 
Fcv bio cv_cottrell
Fcv bio cv_cottrellFcv bio cv_cottrell
Fcv bio cv_cottrell
 
Tech talk
Tech talkTech talk
Tech talk
 
Batch normalization presentation
Batch normalization presentationBatch normalization presentation
Batch normalization presentation
 
Deep Neural Network DNN.docx
Deep Neural Network DNN.docxDeep Neural Network DNN.docx
Deep Neural Network DNN.docx
 
Dynamic Routing Between Capsules.pdf
Dynamic Routing Between Capsules.pdfDynamic Routing Between Capsules.pdf
Dynamic Routing Between Capsules.pdf
 
Neural Architecture Search: Learning How to Learn
Neural Architecture Search: Learning How to LearnNeural Architecture Search: Learning How to Learn
Neural Architecture Search: Learning How to Learn
 
Deep learning from a novice perspective
Deep learning from a novice perspectiveDeep learning from a novice perspective
Deep learning from a novice perspective
 
Machine learning Module-2, 6th Semester Elective
Machine learning Module-2, 6th Semester ElectiveMachine learning Module-2, 6th Semester Elective
Machine learning Module-2, 6th Semester Elective
 
Mnist report ppt
Mnist report pptMnist report ppt
Mnist report ppt
 
Review-image-segmentation-by-deep-learning
Review-image-segmentation-by-deep-learningReview-image-segmentation-by-deep-learning
Review-image-segmentation-by-deep-learning
 
Mnist report
Mnist reportMnist report
Mnist report
 
Cnn
CnnCnn
Cnn
 
Dynamic routing between capsules - A brief presentation
Dynamic routing between capsules - A brief presentationDynamic routing between capsules - A brief presentation
Dynamic routing between capsules - A brief presentation
 
Lalal
LalalLalal
Lalal
 
ML Module 3 Non Linear Learning.pptx
ML Module 3 Non Linear Learning.pptxML Module 3 Non Linear Learning.pptx
ML Module 3 Non Linear Learning.pptx
 
Neural Networks in Data Mining - “An Overview”
Neural Networks  in Data Mining -   “An Overview”Neural Networks  in Data Mining -   “An Overview”
Neural Networks in Data Mining - “An Overview”
 

Mehr von Kyuhwan Jung

(20180524) vuno seminar roc and extension
(20180524) vuno seminar   roc and extension(20180524) vuno seminar   roc and extension
(20180524) vuno seminar roc and extensionKyuhwan Jung
 
(20180715) ksiim gan in medical imaging - vuno - kyuhwan jung
(20180715) ksiim   gan in medical imaging - vuno - kyuhwan jung(20180715) ksiim   gan in medical imaging - vuno - kyuhwan jung
(20180715) ksiim gan in medical imaging - vuno - kyuhwan jungKyuhwan Jung
 
(20180728) kosaim workshop vuno - kyuhwan jung
(20180728) kosaim workshop   vuno - kyuhwan jung(20180728) kosaim workshop   vuno - kyuhwan jung
(20180728) kosaim workshop vuno - kyuhwan jungKyuhwan Jung
 
Generative Adversarial Networks and Their Medical Imaging Applications
Generative Adversarial Networks and Their Medical Imaging ApplicationsGenerative Adversarial Networks and Their Medical Imaging Applications
Generative Adversarial Networks and Their Medical Imaging ApplicationsKyuhwan Jung
 
(2017/06)Practical points of deep learning for medical imaging
(2017/06)Practical points of deep learning for medical imaging(2017/06)Practical points of deep learning for medical imaging
(2017/06)Practical points of deep learning for medical imagingKyuhwan Jung
 
Hello, Recommender System
Hello, Recommender SystemHello, Recommender System
Hello, Recommender SystemKyuhwan Jung
 

Mehr von Kyuhwan Jung (6)

(20180524) vuno seminar roc and extension
(20180524) vuno seminar   roc and extension(20180524) vuno seminar   roc and extension
(20180524) vuno seminar roc and extension
 
(20180715) ksiim gan in medical imaging - vuno - kyuhwan jung
(20180715) ksiim   gan in medical imaging - vuno - kyuhwan jung(20180715) ksiim   gan in medical imaging - vuno - kyuhwan jung
(20180715) ksiim gan in medical imaging - vuno - kyuhwan jung
 
(20180728) kosaim workshop vuno - kyuhwan jung
(20180728) kosaim workshop   vuno - kyuhwan jung(20180728) kosaim workshop   vuno - kyuhwan jung
(20180728) kosaim workshop vuno - kyuhwan jung
 
Generative Adversarial Networks and Their Medical Imaging Applications
Generative Adversarial Networks and Their Medical Imaging ApplicationsGenerative Adversarial Networks and Their Medical Imaging Applications
Generative Adversarial Networks and Their Medical Imaging Applications
 
(2017/06)Practical points of deep learning for medical imaging
(2017/06)Practical points of deep learning for medical imaging(2017/06)Practical points of deep learning for medical imaging
(2017/06)Practical points of deep learning for medical imaging
 
Hello, Recommender System
Hello, Recommender SystemHello, Recommender System
Hello, Recommender System
 

Kürzlich hochgeladen

Generative AI or GenAI technology based PPT
Generative AI or GenAI technology based PPTGenerative AI or GenAI technology based PPT
Generative AI or GenAI technology based PPTbhaskargani46
 
Call Girls Wakad Call Me 7737669865 Budget Friendly No Advance Booking
Call Girls Wakad Call Me 7737669865 Budget Friendly No Advance BookingCall Girls Wakad Call Me 7737669865 Budget Friendly No Advance Booking
Call Girls Wakad Call Me 7737669865 Budget Friendly No Advance Bookingroncy bisnoi
 
Call Girls In Bangalore ☎ 7737669865 🥵 Book Your One night Stand
Call Girls In Bangalore ☎ 7737669865 🥵 Book Your One night StandCall Girls In Bangalore ☎ 7737669865 🥵 Book Your One night Stand
Call Girls In Bangalore ☎ 7737669865 🥵 Book Your One night Standamitlee9823
 
Call for Papers - Educational Administration: Theory and Practice, E-ISSN: 21...
Call for Papers - Educational Administration: Theory and Practice, E-ISSN: 21...Call for Papers - Educational Administration: Theory and Practice, E-ISSN: 21...
Call for Papers - Educational Administration: Theory and Practice, E-ISSN: 21...Christo Ananth
 
Intze Overhead Water Tank Design by Working Stress - IS Method.pdf
Intze Overhead Water Tank  Design by Working Stress - IS Method.pdfIntze Overhead Water Tank  Design by Working Stress - IS Method.pdf
Intze Overhead Water Tank Design by Working Stress - IS Method.pdfSuman Jyoti
 
Double rodded leveling 1 pdf activity 01
Double rodded leveling 1 pdf activity 01Double rodded leveling 1 pdf activity 01
Double rodded leveling 1 pdf activity 01KreezheaRecto
 
Thermal Engineering -unit - III & IV.ppt
Thermal Engineering -unit - III & IV.pptThermal Engineering -unit - III & IV.ppt
Thermal Engineering -unit - III & IV.pptDineshKumar4165
 
data_management_and _data_science_cheat_sheet.pdf
data_management_and _data_science_cheat_sheet.pdfdata_management_and _data_science_cheat_sheet.pdf
data_management_and _data_science_cheat_sheet.pdfJiananWang21
 
Call for Papers - African Journal of Biological Sciences, E-ISSN: 2663-2187, ...
Call for Papers - African Journal of Biological Sciences, E-ISSN: 2663-2187, ...Call for Papers - African Journal of Biological Sciences, E-ISSN: 2663-2187, ...
Call for Papers - African Journal of Biological Sciences, E-ISSN: 2663-2187, ...Christo Ananth
 
CCS335 _ Neural Networks and Deep Learning Laboratory_Lab Complete Record
CCS335 _ Neural Networks and Deep Learning Laboratory_Lab Complete RecordCCS335 _ Neural Networks and Deep Learning Laboratory_Lab Complete Record
CCS335 _ Neural Networks and Deep Learning Laboratory_Lab Complete RecordAsst.prof M.Gokilavani
 
Java Programming :Event Handling(Types of Events)
Java Programming :Event Handling(Types of Events)Java Programming :Event Handling(Types of Events)
Java Programming :Event Handling(Types of Events)simmis5
 
Call Girls Pimpri Chinchwad Call Me 7737669865 Budget Friendly No Advance Boo...
Call Girls Pimpri Chinchwad Call Me 7737669865 Budget Friendly No Advance Boo...Call Girls Pimpri Chinchwad Call Me 7737669865 Budget Friendly No Advance Boo...
Call Girls Pimpri Chinchwad Call Me 7737669865 Budget Friendly No Advance Boo...roncy bisnoi
 
PVC VS. FIBERGLASS (FRP) GRAVITY SEWER - UNI BELL
PVC VS. FIBERGLASS (FRP) GRAVITY SEWER - UNI BELLPVC VS. FIBERGLASS (FRP) GRAVITY SEWER - UNI BELL
PVC VS. FIBERGLASS (FRP) GRAVITY SEWER - UNI BELLManishPatel169454
 
Double Revolving field theory-how the rotor develops torque
Double Revolving field theory-how the rotor develops torqueDouble Revolving field theory-how the rotor develops torque
Double Revolving field theory-how the rotor develops torqueBhangaleSonal
 
ONLINE FOOD ORDER SYSTEM PROJECT REPORT.pdf
ONLINE FOOD ORDER SYSTEM PROJECT REPORT.pdfONLINE FOOD ORDER SYSTEM PROJECT REPORT.pdf
ONLINE FOOD ORDER SYSTEM PROJECT REPORT.pdfKamal Acharya
 
FULL ENJOY Call Girls In Mahipalpur Delhi Contact Us 8377877756
FULL ENJOY Call Girls In Mahipalpur Delhi Contact Us 8377877756FULL ENJOY Call Girls In Mahipalpur Delhi Contact Us 8377877756
FULL ENJOY Call Girls In Mahipalpur Delhi Contact Us 8377877756dollysharma2066
 

Kürzlich hochgeladen (20)

Generative AI or GenAI technology based PPT
Generative AI or GenAI technology based PPTGenerative AI or GenAI technology based PPT
Generative AI or GenAI technology based PPT
 
Water Industry Process Automation & Control Monthly - April 2024
Water Industry Process Automation & Control Monthly - April 2024Water Industry Process Automation & Control Monthly - April 2024
Water Industry Process Automation & Control Monthly - April 2024
 
(INDIRA) Call Girl Bhosari Call Now 8617697112 Bhosari Escorts 24x7
(INDIRA) Call Girl Bhosari Call Now 8617697112 Bhosari Escorts 24x7(INDIRA) Call Girl Bhosari Call Now 8617697112 Bhosari Escorts 24x7
(INDIRA) Call Girl Bhosari Call Now 8617697112 Bhosari Escorts 24x7
 
Call Girls Wakad Call Me 7737669865 Budget Friendly No Advance Booking
Call Girls Wakad Call Me 7737669865 Budget Friendly No Advance BookingCall Girls Wakad Call Me 7737669865 Budget Friendly No Advance Booking
Call Girls Wakad Call Me 7737669865 Budget Friendly No Advance Booking
 
Call Girls In Bangalore ☎ 7737669865 🥵 Book Your One night Stand
Call Girls In Bangalore ☎ 7737669865 🥵 Book Your One night StandCall Girls In Bangalore ☎ 7737669865 🥵 Book Your One night Stand
Call Girls In Bangalore ☎ 7737669865 🥵 Book Your One night Stand
 
Call for Papers - Educational Administration: Theory and Practice, E-ISSN: 21...
Call for Papers - Educational Administration: Theory and Practice, E-ISSN: 21...Call for Papers - Educational Administration: Theory and Practice, E-ISSN: 21...
Call for Papers - Educational Administration: Theory and Practice, E-ISSN: 21...
 
Intze Overhead Water Tank Design by Working Stress - IS Method.pdf
Intze Overhead Water Tank  Design by Working Stress - IS Method.pdfIntze Overhead Water Tank  Design by Working Stress - IS Method.pdf
Intze Overhead Water Tank Design by Working Stress - IS Method.pdf
 
Double rodded leveling 1 pdf activity 01
Double rodded leveling 1 pdf activity 01Double rodded leveling 1 pdf activity 01
Double rodded leveling 1 pdf activity 01
 
Thermal Engineering -unit - III & IV.ppt
Thermal Engineering -unit - III & IV.pptThermal Engineering -unit - III & IV.ppt
Thermal Engineering -unit - III & IV.ppt
 
data_management_and _data_science_cheat_sheet.pdf
data_management_and _data_science_cheat_sheet.pdfdata_management_and _data_science_cheat_sheet.pdf
data_management_and _data_science_cheat_sheet.pdf
 
Call for Papers - African Journal of Biological Sciences, E-ISSN: 2663-2187, ...
Call for Papers - African Journal of Biological Sciences, E-ISSN: 2663-2187, ...Call for Papers - African Journal of Biological Sciences, E-ISSN: 2663-2187, ...
Call for Papers - African Journal of Biological Sciences, E-ISSN: 2663-2187, ...
 
Call Girls in Ramesh Nagar Delhi 💯 Call Us 🔝9953056974 🔝 Escort Service
Call Girls in Ramesh Nagar Delhi 💯 Call Us 🔝9953056974 🔝 Escort ServiceCall Girls in Ramesh Nagar Delhi 💯 Call Us 🔝9953056974 🔝 Escort Service
Call Girls in Ramesh Nagar Delhi 💯 Call Us 🔝9953056974 🔝 Escort Service
 
CCS335 _ Neural Networks and Deep Learning Laboratory_Lab Complete Record
CCS335 _ Neural Networks and Deep Learning Laboratory_Lab Complete RecordCCS335 _ Neural Networks and Deep Learning Laboratory_Lab Complete Record
CCS335 _ Neural Networks and Deep Learning Laboratory_Lab Complete Record
 
Java Programming :Event Handling(Types of Events)
Java Programming :Event Handling(Types of Events)Java Programming :Event Handling(Types of Events)
Java Programming :Event Handling(Types of Events)
 
Call Girls Pimpri Chinchwad Call Me 7737669865 Budget Friendly No Advance Boo...
Call Girls Pimpri Chinchwad Call Me 7737669865 Budget Friendly No Advance Boo...Call Girls Pimpri Chinchwad Call Me 7737669865 Budget Friendly No Advance Boo...
Call Girls Pimpri Chinchwad Call Me 7737669865 Budget Friendly No Advance Boo...
 
PVC VS. FIBERGLASS (FRP) GRAVITY SEWER - UNI BELL
PVC VS. FIBERGLASS (FRP) GRAVITY SEWER - UNI BELLPVC VS. FIBERGLASS (FRP) GRAVITY SEWER - UNI BELL
PVC VS. FIBERGLASS (FRP) GRAVITY SEWER - UNI BELL
 
Double Revolving field theory-how the rotor develops torque
Double Revolving field theory-how the rotor develops torqueDouble Revolving field theory-how the rotor develops torque
Double Revolving field theory-how the rotor develops torque
 
ONLINE FOOD ORDER SYSTEM PROJECT REPORT.pdf
ONLINE FOOD ORDER SYSTEM PROJECT REPORT.pdfONLINE FOOD ORDER SYSTEM PROJECT REPORT.pdf
ONLINE FOOD ORDER SYSTEM PROJECT REPORT.pdf
 
Roadmap to Membership of RICS - Pathways and Routes
Roadmap to Membership of RICS - Pathways and RoutesRoadmap to Membership of RICS - Pathways and Routes
Roadmap to Membership of RICS - Pathways and Routes
 
FULL ENJOY Call Girls In Mahipalpur Delhi Contact Us 8377877756
FULL ENJOY Call Girls In Mahipalpur Delhi Contact Us 8377877756FULL ENJOY Call Girls In Mahipalpur Delhi Contact Us 8377877756
FULL ENJOY Call Girls In Mahipalpur Delhi Contact Us 8377877756
 

Dynamic Routing Between Capsules

  • 1.
  • 3. Stack of Layers with Convolution, Subsampling and Nonlinearity Operations Modern CNNs  Convolution Layer –Filtering of unimportant information and extraction of salient local feature.  Subsampling Layer –Introduction of local transition invariance, reduction of computation and enlargement of receptive field.  Nonlinearity –Increase of capacity to model more complex input-output relationship.
  • 4. Mistake and Disaster What Geoff. Hinton Says “The pooling operation used in convolutional neural networks is a big mistake and the fact that it works so well is a disaster.” “Internal data representation of a convolutional neural network does not take into account important spatial hierarchies between simple and complex objects.”
  • 5.  What our brain does when we see the world is ‘inverse graphics’ –Extraction of hierarchical representation of the visual information –The representation is ‘equivariance’, not ‘invariance’ to changes in viewpoint Computer Graphics vs Computer Vision Inverse Graphical Approach
  • 6.  Sub-sampling tries to make the neural activities invariant for small changes in viewpoint. –This is a silly goal, motivated by the fact that the final label needs to be viewpoint-invariant.  Its better to aim for equivariance: Changes in viewpoint lead to corresponding changes in neural activities. –In the perceptual system, its the weights that code viewpoint-invariant knowledge, not the neural activities. Better Representation via Better Objective Invariance vs Equivariance
  • 8.  Transitional invariance –The output is unchanged by small shift of object or parts  Reduction of input –Allowing more types of feature in the next layer  Bad fit to psychology of shape perception –We have intrinsic coordinate frames to objects and use it while recognizing things.  Solution to wrong problem –We need equivariance not invariance. Disentangling rather than discarding.  Fails to use underlying linear structure –We need to make used of the natural linear manifold which will help us to extrapolate.  Poor wary to do dynamic routing –We need to route each part of the input to the neurons that know how to deal with it. Routing via most active neuron is primitive. Pros and Cons Arguments on Pooling
  • 9.  If a low-level part moves to a very different position it will be represented by a different capsule –‘Place-coded’ equivariance  If a part only moves a small distance it will be represented by the same capsule but the pose of outputs of the capsule will change –‘Rate-coded’ equivariance  Higher-level capsules have bigger domains so low-level ‘place-coded’ equivariance gets converted into high-level ‘rate-coded’ equivariance.  Without sub-sampling, CNNs can give ‘place-coded’ equivariance for discrete translations. Place-coding vs Rate-coding Two types of equivariance Representation Image Translated Representation Translated Image
  • 10.  Neuron –The activation of a specific neuron = the likelihood of specific feature in the input image. –We need specialized mechanism(e.g. generative models) to make each neuron represent parameters of feature. Neuron vs Capsule  Capsule – A group of neurons which not only capture the likelihood of specific features but also the parameters related to the features. –The magnitude of activity vector represents the probability of detecting specific features and its orientation represents its parameters or properties (position, orientation, scale, deformation, color, hue, texture …)
  • 11.  Neuron –The inputs of lower layers and the outputs of upper layers are scalars. –After aggregating inputs multiplied by weights, nonlinearity function such as ReLU is applied. Neuron vs Capsule  Capsule – The inputs of the lower layers and the outputs of upper layers are vectors. – Instead of using nonlinearity function such as ReLU, we used squashing function.
  • 12.  Prediction vector –With the previous capsule output and transformation matrix , we computer prediction vector as  The capsule output of next layer –Then the capsule output of the next layer is computed as , where  Coupling coefficient –Here, are called coupling coefficient which is trained with dynamic routing algorithm. –We impose restriction that = 1 which is achieved by softmax function of relevancy or similarity score which is initialized with zeros and progressively updated as follows(which reminds me of Hebbian learning rule): Participants in the Algorithm Dynamic Routing Algorithm
  • 13.  Routing Algorithm –The pseudo-code of routing algorithm is as follows:  Loss Function –For each digit capsule , the loss function is margin loss as where when digit is present and . The default Routing Algorithm and Loss Function Dynamic Routing Algorithm
  • 14.  Model Architecture  Reconstruction Regularization –To encourage the digit capsule to encode instantiation parameters of the input digits, we add reconstruction loss (with weight 0.0005) to margin loss using a decoder with 3 fully connected layers. CapsNet for MNIST The Architecture of Capsule Network
  • 15.  Classification + Reconstruction in One Diagram CapsNet for MNIST The Architecture of Capsule Network
  • 17.  Prediction and Reconstruction Example – = label, prediction, reconstruction target  Classification Accuracy MNIST Dataset Hand-written Digit Classification
  • 18.  Improvement of recognition performance on novel viewpoints smallNORB Dataset Object Recognition from Various Angle
  • 20.  We can train better models with less training data which is quite beneficial for medical image analysis where the annotations are scarce and expensive to get.  We need to study more on how our brain works. Pure data-driven approach with overcomplex models might be wasting some resource(both data and computations)  The CapsNet is the combination of credit assignment mechanism of CNN and instantization parameter learning of generative models.  CapsNet is slower than CNN counterpart according to the routing update iterations.  CpasNet has not yet proven its effectiveness in large-scale visual recognition problems such as ‘ImageNet’  Each element in the activity vector does not always represent meaningful properties of input image  Cosine similarity may not be the best measure of similarity between the predicted capsule output and actual capsule output. Departure from ‘Conventional Neural Networks’ Discussion
  • 21.  Paper – Original NIPS 2017 Paper : https://arxiv.org/abs/1710.09829 – ICLR 2018 Paper : https://openreview.net/pdf?id=HJWLfGWRb  Video – Geofrrey Hinton's talk - "What is wrong with convolutional neural nets?" • https://www.youtube.com/watch?v=rTawFwUvnLE&feature=youtu.be – Siraj Raval lecture - "Capsule Networks : An Improvement to Convolutional Networks" • https://www.youtube.com/watch?v=VKoLGnq15RM&feature=youtu.be&t=13m59s  Related News and Blog Posts – https://medium.com/@pechyonkin/understanding-hintons-capsule-networks-part-i-intuition-b4b559d1159b – https://hackernoon.com/what-is-a-capsnet-or-capsule-network-2bfbe48769cc – https://www.nextobe.com/single-post/2017/11/02/CapsNetCapsule-Network – https://www.wired.com/story/googles-ai-wizard-unveils-a-new-twist-on-neural-networks/ – https://jhui.github.io/2017/11/03/Dynamic-Routing-Between-Capsules/  Implementation – https://github.com/nishnik/CapsNet-PyTorch – https://github.com/debarko/CapsNet-Tensorflow – https://github.com/XifengGuo/CapsNet-Keras – https://github.com/jaesik817/adv_attack_capsnet References
  • 22. hello@vuno.co Putting the world’s medical data to work

Hinweis der Redaktion

  1. 아산병원 2700병상, 유동인구 5만
  2. https://hackernoon.com/what-is-a-capsnet-or-capsule-network-2bfbe48769cc
  3. https://medium.com/@pechyonkin/understanding-hintons-capsule-networks-part-i-intuition-b4b559d1159b https://hackernoon.com/what-is-a-capsnet-or-capsule-network-2bfbe48769cc
  4. http://aboutintelligence.blogspot.kr/2009/01/computer-vision-vs-computer-graphics.html https://www.slideshare.net/horizongir/sprite-rendering 그러니까 CNN은 Translation이나 Rotation에 Feature가 Invariant하도록 강제하는 것인데 이것이 자연스러운 것이냐? 이렇게 하기 위해 너무 많은 데이터가 필요하다. (모든 각도에서 바라본 자동차 사진) Sub-sampling tries to make the neural activities invariant for small changes in viewpoint. – This is a silly goal, motivated by the fact that the final label needs to be viewpoint-invariant. • Its better to aim for equivariance: Changes in viewpoint lead to corresponding changes in neural activities. – In the perceptual system, its the weights that code viewpoint-invariant knowledge, not the neural activities.
  5. http://cseweb.ucsd.edu/~gary/cs200/s12/Hinton.pdf https://medium.com/@pechyonkin/understanding-hintons-capsule-networks-part-i-intuition-b4b559d1159b
  6. 아산병원 2700병상, 유동인구 5만
  7. Part의 큰 변화는 다른 Object를 의미할 수 있음. Part의 작은 변화는 Pose의 변화일 수 있음. 따라서 Low-level Place coding의 변화는 Rate를 바꿀 수 있음.
  8. 아산병원 2700병상, 유동인구 5만
  9. https://openreview.net/pdf?id=HJWLfGWRb
  10. 아산병원 2700병상, 유동인구 5만