Publications

Detailed Information

Deep Neural Network Latent Space Analysis using Decision Boundary and Attention Style CapsuleNet : 결정 경계와 어텐션 캡슐 네트워크를 이용한 잠재공간 특성백터 분석

DC Field Value Language
dc.contributor.advisor강명주-
dc.contributor.author서현-
dc.date.accessioned2022-06-08T06:36:35Z-
dc.date.available2022-06-08T06:36:35Z-
dc.date.issued2022-
dc.identifier.other000000169513-
dc.identifier.urihttps://hdl.handle.net/10371/181126-
dc.identifier.urihttps://dcollection.snu.ac.kr/common/orgView/000000169513ko_KR
dc.description학위논문(박사) -- 서울대학교대학원 : 자연과학대학 협동과정 계산과학전공, 2022.2. 강명주.-
dc.description.abstractWhile the deep learning model produces overwhelming performance in many domains, it is not known what latent space the deep learning model embedding, what features it learns, and how it separates features.
An accurate understanding of the learning process of deep learning is not perfect until now and is still an open problem.

In this thesis, we try to broaden our understanding of the latent space of the deep neural network in two ways.
In the first chapter, we experimentally investigate the relationships with the vision boundary in the latent space of the deep neural network through several toy experiments.
The decision boundary is obtained by using and adversarial attack methods in the latent space where deep neural network embeds.
We analyze the relationship between the decision boundary and the latent space manual obtained by perturbing the image.

In the second chapter, the characteristics of the latent space is examined by constraining a network architecture design.
We propose a new network module called an attention-style capsulenet with improved version of capsulenet.
The value of each capsule is perturbed to determine which image feature is trained by the deep neural network.
-
dc.description.abstract딥러닝 모델이 많은 도메인에서 압도적인 성능을 내는데 반해 딥러닝 모델이 어떤 잠재 공간을 만드는지, 어떤 특징를 학습 하는지 정확히 알려지지 않고 있다.
딥러닝의 학습 과정에 대한 정확한 이해는 현재까지 완벽하지 않고 열린 문제이다.


이번 연구에서는 크게 2가지 방법으로 딥러닝이 학습한 잠재공간에 대해 이해를 넓히고자 하였다.
첫번째 챕터에서는 딥러닝 모델의 잠재공간에서 결정 경계를 이용하여 잠재공간의 특성을 여러 실험들을 통해 분석하였다.
적대적 공격 방법을 이용해서 결정경계 벡터를 구하고 인풋에 노이즈를 추가하여 잠재공간의 다양체 벡터를 구하였다.
결정경계 벡터와 잠재공간의 다양체 벡터 사이의 관계를 통해 잠재공간의 다양체의 구성을 분석하였다.

두번째 챕터에서는 캡슐이라는 특수한 설계로서 잠재공간을 제한하여 학습된 잠재공간의 특징을 살펴보았다.
네트워크 구조는 캡슐넷을 개선한 어텐션 스타일의 캡슐넷을 제안하고, 각 캡슐들을 값을 변동시켜 딥러닝 모델이 실제 이미지공간의 어떤 특징을 분석하여 캡슐 잠재공간으로 매핑하는지 확인하고 인풋 데이터에서의 변형에 대한 캡슐값의 변동을 분석하였다.
-
dc.description.tableofcontents1 Introduction 1
2 Relation between Feature Manifold and Decision Boundary 4
2.1 Related Work 7
2.1.1 Manifold Hypothesis 7
2.1.2 Manifold Learning Methods 8
2.1.3 Adversarial Attack 12
2.1.4 Explain AI (Visualization methods) 14
2.2 Distribution of angles between latent manifold and the decision boundary 16
2.2.1 Experiment detail 16
2.2.2 Experiment results 18
2.3 Near-local manifold curvature 23
2.3.1 Experiment detail 23
2.3.2 Experiment results 23
2.4 Miscellaneous experiments 28
2.4.1 Does adversarial attack really mean a vulnerability in deep learning models 28
2.4.2 Is the manifold's shape related to the performance of the model 30
3 Attention style Capsulenet 32
3.1 Related Works 36
3.2 Proposed Method 39
3.2.1 Primary Caps Layer 40
3.2.2 Capsule Activation 40
3.2.3 Conv Caps Layer 41
3.2.4 Fully Conv Caps Layer 44
3.2.5 Margin Loss and Reconstruction Regularizer 44
3.3 Experiments 46
3.3.1 Classiffication Results on MNIST and affNIST 47
3.3.2 Classiffication Results on CIFAR-10 49
3.3.3 Robustness to hyperparameters 51
3.3.4 Transformation Equivariance 52
4 Conclusion and Future Works 57
-
dc.format.extentviii, 66-
dc.language.isokor-
dc.publisher서울대학교 대학원-
dc.subjectDeep Learning-
dc.subjectLatent Manifold-
dc.subjectAdversarial Attack-
dc.subjectDecision Boundary-
dc.subjectCapsuleNet-
dc.subject.ddc004-
dc.titleDeep Neural Network Latent Space Analysis using Decision Boundary and Attention Style CapsuleNet-
dc.title.alternative결정 경계와 어텐션 캡슐 네트워크를 이용한 잠재공간 특성백터 분석-
dc.typeThesis-
dc.typeDissertation-
dc.contributor.AlternativeAuthorSeo Hyun-
dc.contributor.department자연과학대학 협동과정 계산과학전공-
dc.description.degree박사-
dc.date.awarded2022-02-
dc.identifier.uciI804:11032-000000169513-
dc.identifier.holdings000000000047▲000000000054▲000000169513▲-
Appears in Collections:
Files in This Item:

Altmetrics

Item View & Download Count

  • mendeley

Items in S-Space are protected by copyright, with all rights reserved, unless otherwise indicated.

Share