Detail View
Subject-Adaptive Learning for Personalized EEG Analysis against Inter-Subject Variability
WEB OF SCIENCE
SCOPUS
- Title
- Subject-Adaptive Learning for Personalized EEG Analysis against Inter-Subject Variability
- Alternative Title
- 피험자 간 변동성을 극복하는 개인화 뇌파 분석을 위한 적응형 학습
- DGIST Authors
- Sion An ; Sang Hyun Park ; Li Shen
- Advisor
- 박상현
- Co-Advisor(s)
- Li Shen
- Issued Date
- 2026
- Awarded Date
- 2026-02-01
- Type
- Thesis
- Description
- Electroencephalography, Meta-learning, Few-shot learning, Subject-adaptive learning, Foundation model
- Abstract
-
Advancements in deep learning have innovated electroencephalography (EEG) based brain-computer interfaces (BCI). However, their practical application is severely lim- ited by high inter-subject and inter-dataset variability, which hinders the personalization of the models. To achieve the personalized BCI, this thesis aims to develop models by reducing the data dependency required for subject adaptation. First, to mitigate the extensive data collection burden for new subjects, we propose a few-shot learning frame- work. By leveraging a dual attention relation network and a fine-tuning strategy, our model effectively analyze EEG signals of new subjects using only a few labeled signals. To further advance data efficiency, we then eliminate the need for task-specific EEG sig- nals by adapting the model using only resting-state (RS) EEG signals. We introduce a novel method that fuses subject information from RS EEG signals with general task knowledge within a gradient-based meta-learning framework, enabling adaptation with- out active user participation. Lastly, to build a universal model, we introduce an EEG foundation model that generalizes across diverse subjects and tasks. By employing a fea- ture disentanglement mechanism within a meta-learning framework, the model explicitly separates subject- and task-dependent representations. This approach results in a highly adaptable model that achieves state-of-the-art performance on new subjects and tasks. The diverse methodologies proposed in this study are expected to significantly enhance the practical applicability of artificial intelligence in EEG analysis.|딥러닝 기술의 발전은 뇌파 기반 뇌-컴퓨터 인터페이스 분야에 혁신을 가져왔습니다. 하지만 높은 개인 간, 데이터셋 간의 편차는 모델의 개인화를 저해하여 적용을 크게 제한합니다. 본 논문은 개인화된 뇌-컴퓨터 인터페이스를 구현하기 위해 피험자 적응에 필요한 데이터 의존도를 줄이는 모델 개발을 목표로 합니다.
더보기
첫째, 새로운 피험자에 대한 광범위한 데이터 수집 부담을 완화하기 위해 소량 데이터 학습 프레임워크를 제안합니다. 이중 주의 관계망과 미세 조정 전략을 활용하여 제안 모델은 소수의 라벨이 있는 신호만으로 새로운 피험자의 뇌파 신호를 효과적으로 분석합니다.
둘째, 데이터 효율성을 더욱 높이기 위해 과제 수행과 무관한 휴지기 상태 뇌파 신호만을 사용하여 모델을 적응시킴으로써 과제 관련 뇌파 신호의 필요성을 제거합니다. 경사도 기반 메타 학습 프레임워크 내에서 휴지기 상태 뇌파 신호로부터 얻은 피험자 정보와 일반적인 과제 지식을 융합하는 새로운 방법을 도입하여 사용자의 적극적인 참여 없이도 모델 적응이 가능하도록 합니다.
마지막으로, 범용 모델을 구축하기 위해 다양한 피험자와 과제에 걸쳐 일반화되는 뇌파 파운데이션 모델을 소개합니다. 메타 학습 프레임워크 내에서 특징 분리 메커니즘을 사용하여 모델은 피험자 의존적 표현과 과제 의존적 표현을 명시적으로 분리합니다. 이 접근법은 새로운 피험자와 과제에 대해 최고 수준의 성능을 달성하는 높은 적응성을 가진 모델을 만듭니다. 본 연구에서 제안된 다양한 방법론들은 뇌파 분석 분야에서 인공지능의 실용성을 크게 향상시킬 것으로 기대됩니다.
- Table Of Contents
-
Contents i
List of Tables iv
List of Figures vi
I. Introduction 1
1 Background and Motivations 1
2 Contributions and Outline 2
3 Publications 3
3.1 Excluded Research 4
II. Few-Shot Learning for EEG signal Classification 6
1 Introduction 6
2 Related works 8
3 Method 11
3.1 Embedding Module 12
3.2 Temporal-Attention Module 14
3.3 Aggregation-Attention Module 15
3.4 Relation Module 15
3.5 Fine-Tuning Strategy 16
4 Experimental Settings 17
4.1 Dataset 17
4.2 Validation 17
4.3 Comparison 18
5 Results and Discussion 20
5.1 Cross-subject Evaluation on BCI-2b 20
5.2 Cross-dataset Evaluation on BCI-2a 22
5.3 Validation on GIST 24
5.4 Attention score visualization 26
5.5 Performance Differences Between Subjects 27
5.6 Robustness with respect to Support set Variations 28
5.7 Effect of Embedding Module Variations 30
6 Conclusion 31
III. Subject-Adaptive Learning using Resting-State EEG Signal 32
1 Introduction 33
2 Related works 35
2.1 EEG Classification using Deep Learning 35
2.2 Mitigation of Inter-Subject Variability 36
2.3 Meta-Learning 37
2.4 Data-free synthesis 38
2.5 Efficient Fine-tuning Methods 38
3 Methodology 39
3.1 Problem settings 39
3.2 Overview 39
3.3 Model architecture 40
3.4 Loss functions 41
3.5 Fusion of resting-state EEG signal with task-specific information 42
3.6 ResTL: Subject-adaptive transfer learning 42
3.7 ResML: Subject-adaptive meta-learning 44
4 Experimental Settings 47
4.1 Dataset 47
4.2 Experimental details 48
4.3 Comparison methods 48
5 Experimental Results 49
5.1 Comparison of classification results 49
5.2 Effect of RS EEG signals and subject-dependent features 50
5.3 Effect of loss functions 54
5.4 Distribution of fused signals 54
5.5 Effect of the number of RS EEG signals 55
5.6 Effect of the adapter 55
5.7 Effect of the adaptation epochs 56
5.8 Task generalization and extendability 56
6 Discussion and Conclusion 57
IV. Meta-Learning based EEG Foundation Model 58
1 Introduction 58
2 Method 60
2.1 Problem Formulation 61
2.2 Patch Embedding 61
2.3 Feature Disentanglement Architecture 62
2.4 Supervised Meta-Learning 62
2.5 Self-Supervised Contrastive Learning 64
2.6 Pre-training and Fine-tuning 64
3 Experimental Settings 65
3.1 Pre-training Datasets 65
3.2 Downstream Datasets 65
3.3 Comparison Methods 66
3.4 Evaluation Metrics 67
3.5 Implementation Details 67
4 Results 67
4.1 Motor Imagery Classification 67
4.2 Generalization to Unseen Downstream Tasks 68
4.3 Few-Shot Learning Performance 70
4.4 Ablation Study on Loss Functions 70
4.5 Ablation Study on Functional-Embedding 72
4.6 Functional Connectivity Analysis 72
5 Conclusion 74
V. Concluding Remarks 75
1 Conclusion 75
2 Future Work 75
VI. Acknowledgement 77
References 78
VII. 요약문 95
- URI
-
https://scholar.dgist.ac.kr/handle/20.500.11750/59611
http://dgist.dcollection.net/common/orgView/200000942204
- Degree
- Doctor
- Publisher
- DGIST
File Downloads
- There are no files associated with this item.
공유
Total Views & Downloads
???jsp.display-item.statistics.view???: , ???jsp.display-item.statistics.download???:
