Publications

Detailed Information

Information Theoretic Analysis of Machine Learning : 기계 학습의 정보 이론적 해석

DC Field Value Language
dc.contributor.advisor김석-
dc.contributor.author이성엽-
dc.date.accessioned2023-12-06T01:23:15Z-
dc.date.available2023-12-06T01:23:15Z-
dc.date.issued2022-
dc.identifier.other000000171076-
dc.identifier.urihttps://hdl.handle.net/10371/197609-
dc.identifier.urihttps://dcollection.snu.ac.kr/common/orgView/000000171076ko_KR
dc.description학위논문(박사) -- 서울대학교대학원 : 자연과학대학 물리·천문학부(물리학전공), 2022. 2. 김석.-
dc.description.abstractI aim to provide a deeper understanding of how machine learning works so
great for solving various problems. Using information theory, I study the information
flow, internal representations, and parameter optimization of neural
networks. First, I visualize the compression and transmission of information
flows of various types of autoencoders, and examine how the various models
remove irrelevant information to reproduce input data. Second, I observe that
the internal representation of neural networks is so special that the frequency
of distinct representations follows scale-invariant power laws both in supervised
and unsupervised learning. I derive how this universal behavior can naturally
arise without explicit regularization during the learning process. Finally, I introduce
the mirror descent algorithm in terms of information geometry, and
explain how the learning algorithm can effectively update the parameters of
learning models in the dual space of the primary parameter space. In conclusion,
information theory and geometry are excellent tools to visualize, analyze,
and optimize neural networks.
-
dc.description.abstract이 논문은 기계 학습이 다양한 문제들을 효과적으로 해결하는 원리에 대해 심층
적으로 이해하는 것을 목표로 하며 정보 이론을 활용하여 신경망의 정보 흐름,
내부 표현, 변수 최적화를 연구한다. 먼저, 다양한 종류의 오토인코더에서 압축과
전송으로 이루어지는 정보 흐름을 시각화하고 각 모형들이 입력 데이터 재생성과
무관한 정보를 어떻게 제거하는지 알아본다. 둘째로, 지도 학습과 비지도 학습에서
신경망의 내부 표현이 크기 불변 멱법칙을 따르는 현상을 관찰하고 학습 과정에서
명시적인 규제 없이 어떻게 이 현상이 보편적으로 나타나는지 유도한다. 끝으로,
거울 하강 알고리즘을 정보 기하학 관점에서 소개하고 변수 공간의 쌍대 공간을
통해 학습 변수들이 어떻게 효과적으로 갱신되는지 알아본다. 결과적으로 정보
이론과 정보 기하학이 신경망 내부를 시각화하고, 분석하고, 최적화하는데 유용한
기법임을 확인한다.
-
dc.description.tableofcontentsAbstract i
Chapter 1 Introduction 1
Chapter 2 Information flows of machine learning 5
2.1 Information plane analysis 5
2.2 Representation learning in autoencoders 8
2.2.1 Information plane of autoencoders 8
2.2.2 Various types of autoencoders 9
2.3 Estimation of mutual information 13
2.4 Information plane of autoencoders 17
2.4.1 Vanilla autoencoders 17
2.4.2 Sparse activity and constrained weights 20
2.4.3 Constrained latent space 22
2.5 Conclusion 25
Chapter 3 Scale-invariant representation of machine learning 28
3.1 Internal representation of machine learning 28
3.2 Power laws in machine learning 30
3.2.1 Unsupervised learning 30
3.2.2 Authentic supervised learning 33
3.2.3 Self-supervised learning 34
3.2.4 Power laws with data distribution 36
3.3 Theoretical analysis of power laws 39
3.3.1 Resolution and relevance 42
3.3.2 Unsupervised learning 43
3.3.3 Supervised learning 44
3.3.4 Self-supervised learning 45
3.3.5 Authentic supervised learning 46
3.4 Conclusion 49
Chapter 4 Mirror descent: learning via dual geometry 51
4.1 Optimization for machine learning 51
4.2 Information Geometry 53
4.3 Mirror descent 56
4.4 Experimental results of mirror descent 61
4.5 Additional results of mirror descent 64
4.6 Analysis of mirror descent 71
4.7 Conclusion 72
Chapter A Matrix-based kernel estimator of mutual information 74
Chapter B Manifold learning of label autoencoder 80
Bibliography 82
초록 94
-
dc.format.extent96-
dc.language.isoeng-
dc.publisher서울대학교 대학원-
dc.subjectMachine learning-
dc.subjectDeep learning-
dc.subjectInformation theory-
dc.subjectInformation geometry-
dc.subjectStatistical physics-
dc.subject.ddc523.01-
dc.titleInformation Theoretic Analysis of Machine Learning-
dc.title.alternative기계 학습의 정보 이론적 해석-
dc.typeThesis-
dc.typeDissertation-
dc.contributor.AlternativeAuthorLee Sung Yeop-
dc.contributor.department자연과학대학 물리·천문학부(물리학전공)-
dc.description.degreeDoctor-
dc.date.awarded2022-02-
dc.identifier.uciI804:11032-000000171076-
dc.identifier.holdings000000000047▲000000000054▲000000171076▲-
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