권호기사보기
| 기사명 | 저자명 | 페이지 | 원문 | 기사목차 |
|---|
결과 내 검색
동의어 포함
Title Page 2
Abstract 6
초록 8
Contents 12
Chapter 1. Introduction 22
1.1. Background 22
1.2. Motivation 28
1.3. Research goals 30
1.4. Scope of this thesis 30
1.5. Contributions 31
1.6. The organization of this thesis 32
Chapter 2. Preliminary and literature review 34
2.1. Human Activity Recognition (HAR) 34
2.1.1. Vision-based HAR 34
2.1.2. Sensor-based HAR 35
2.1.3. Type of sensors for human activity recognition 37
2.1.4. Pipeline of sensor-based HAR dataset creation 39
2.2. Federated learning 42
2.2.1. Statistical heterogeneity 47
2.2.2. Participation heterogeneity 48
2.2.3. Data quality heterogeneity (noisy label) 49
Chapter 3. Model Contrastive Learning and Adaptive Control Variates 51
3.1. Problem definition 51
3.2. Methodology 53
3.2.1. Model architecture 53
3.2.2. Local training phase 54
3.2.3. Local parameter update phase 55
3.2.4. FL algorithm 57
3.3. Experiments 60
3.3.1. Dataset 60
3.3.2. Experimental setup 62
3.3.3. Performance comparison 63
3.3.4. Convergence Speed Comparison 64
3.3.5. Handling real-world HAR skewness 66
3.3.6. Visualization of extracted embeddings 70
3.4. Discussion 70
Chapter 4. Common and personalized model 74
4.1. Problem definition 74
4.2. Methodology 75
4.2.1. Common model 75
4.2.2. The personalized model 77
4.2.3. The fusion layer 77
4.2.4. PerCom FL algorithm 79
4.3. Experiments 80
4.3.1. Baselines 80
4.3.2. Implementation details 83
4.3.3. Comparison with other baselines 84
4.3.4. Communication efficiency comparison 87
4.3.5. T-SNE inspection 88
4.4. Discussion 89
Chapter 5. Handling noisy label 92
5.1. Problem definition 92
5.2. Methodology 93
5.2.1. Model structure 94
5.2.2. Noisy client detection 100
5.2.3. Client training 102
5.2.4. Update aggregation 108
5.3. Experiments 109
5.3.1. Experiment setup 110
5.3.2. Experiment variation 111
5.3.3. Evaluation at low percentage of noisy clients (Γ= 0.4) 112
5.3.4. Evaluation at high percentage of noisy clients (Γ= 0.7) 114
5.3.5. Training convergence comparison 117
5.3.6. Noisy client contribution 120
5.3.7. Label accuracy correction 122
5.3.8. Computation time comparison 123
5.4. Discussion 126
Chapter 6. Conclusion and Future Work 129
6.1. Conclusion 129
6.2. Future work 131
Bibliography 133
Figure 1. Organization of this thesis 33
Figure 2. Sensor-based HAR pipeline for classification task 40
Figure 3. Typical federated learning workflow 43
Figure 4. Federated learning category based on data distribution (a)... 45
Figure 5. Framework of FedCoad 53
Figure 6. Class distribution in the (a) HHAR (b) MotionSense, and (c)... 61
Figure 7. FL baselines F1-score convergence graph (a) MotionSense (b)... 65
Figure 8. Distribution of data among parties for different skew scenarios:... 67
Figure 9. T-SNE plots for the Motionsense dataset comparing (a)... 71
Figure 10. T-SNE plots for the WISDM dataset comparing (a)... 71
Figure 11. Framework of the proposed framework 76
Figure 12. Fusion mechanism inner working of PerCom 78
Figure 13. Test accuracy plot during training 85
Figure 14. TSNE visualization of the common model features 90
Figure 15. Overview of FedNoel framework 94
Figure 16. Model structure that includes label embedding 99
Figure 17. A Feature encoder and a classifier for classifying activities 111
Figure 18. Convergence of FL baselines on USC-HAD 118
Figure 19. Convergence of FL baselines on PAMAP 119
Figure 20. The noisy client's average label accuracy progression of the... 124
Figure 21. The noisy client's average label accuracy progression of the... 125
*표시는 필수 입력사항입니다.
| 전화번호 |
|---|
| 기사명 | 저자명 | 페이지 | 원문 | 기사목차 |
|---|
| 번호 | 발행일자 | 권호명 | 제본정보 | 자료실 | 원문 | 신청 페이지 |
|---|
도서위치안내: / 서가번호:
우편복사 목록담기를 완료하였습니다.
*표시는 필수 입력사항입니다.
저장 되었습니다.