Toward fair and robust deep neural networks under spurious correlations

다국어 초록 (Multilingual Abstract)

Spurious correlations in deep neural networks (DNNs) pose significant challenges in supervised classification, leading to biased predictions and poor generalization in real-world scenarios. This dissertation addresses these challenges through the lens of three critical problems: group fairness, group robustness, and bias transfer, all within the context of classification tasks. Each issue is explored in distinct domains, namely natural image classification, medical diagnosis using brain imaging, and continual learning.

First, we address group fairness in natural image classification, focusing on mitigating performance disparities among demographic groups. We introduce a fairness- aware feature distillation method that leverages Maximum Mean Discrepancy (MMD) to train fair models while maintaining predictive accuracy. Extensive experiments on synthetic and real-world datasets, including CelebA, demonstrate the effectiveness of the proposed approach in achieving fairness without sacrificing performance.

Next, we investigate group robustness in medical diagnosis, particularly in predicting the progression of mild cognitive impairment (MCI) to Alzheimer’s disease (AD) using structural MRI data. By defining groups based on diagnosis labels and age, we identify significant performance disparities among subpopulations. To address these issues, we propose a novel framework that incorporates adaptive modulation layers and a decoupled classifier thereby improving the performance of underrepresented or underperforming groups. Evaluation on real-world datasets, such as ADNI and NACC, highlights significant improvements in group robustness.

Finally, we explore the problem of bias transfer in continual learning (CL), where spurious correlations learned in earlier tasks may persist or even distort previously acquired knowledge through backward transfer. To tackle this, we propose a bias-aware CL framework along with a method that employs group-class balanced sampling. Experiments conducted within this framework demonstrate that our approach effectively mitigates model bias caused by bias transfer in CL, while maintaining robust performance across sequential tasks. These results highlight its potential as a simple yet effective baseline.

Overall, this dissertation advances the understanding of fairness, robustness, and bias transfer in classification tasks using DNNs under spurious correlations, and proposes practical solutions applicable across various domains. These contributions provide a foundation for future research in developing equitable, robust, and adaptive machine learning systems.

국문 초록 (Abstract)

심층 신경망에서의 허위 상관은 지도 학습 기반 분류 문제에서 주요한 도전 과제로 작용하며, 현실 세계에서 예측 편향과 일반화 성능 저하를 초래한다. 본 논문은 이러 한 문제를 집단 공정성, 집단 강건성, 편향 전이라는 세 가지 핵심 과제를 중심으로 다룬다. 자연 이미지 분류, 뇌 영상 기반 진단, 연속 학습 등 서로 다른 응용 환경에서 각 과제를 분석하고 이에 대한 해결책을 제시한다.

먼저, 자연 이미지 분류에서 인구통계학적 집단 간 성능 격차를 완화하기 위해, 최대 평균 불일치 (Maximum Mean Discrepancy, MMD) 를 활용한 공정성 인식 특성 증류 방법을 제안한다. CelebA를 포함한 다양한 합성 및 실제 데이터셋을 대상으로 한 실험을 통해, 본 방법이 예측 성능을 유지하면서도 공정성을 효과적으로 향상시킴을 확인하였다.

다음으로, 구조적 뇌 MRI 데이터를 활용해 경도 인지 장애 (MCI) 에서 알츠하이머병 (AD) 으로의 진행을 예측하는 의료 진단 문제에서 집단 강건성을 살펴본다. 진단 결과와 연령에 따라 집단을 구성한 뒤, 집단 간 성능 차이가 뚜렷하게 나타남을 확인하였다. 이를 해결하기 위해 적응형 변조 계층과 분리된 분류기로 구성된 새로운 프레임워크를 제안하였으며, ADNI 및 NACC 데이터셋을 통해 해당 방법이 집단 강건 성을 크게 개선함을 입증하였다.

마지막으로, 연속 학습에서 발생하는 편향 전이 문제를 다룬다. 초기 과제에서 학습된 허위 상관이 이후 과제에 영향을 미치거나, 이전에 습득한 지식을 왜곡하는 역방향 전이가 발생할 수 있다. 이를 분석 및 해결하기 위해, 편향 인식 연속 학습 프레임워크와 집단-클래스 균형 샘플링 기반의 방법을 제안하였다. 제안한 프레임 워크를 기반으로 한 실험 결과, 제안된 방법은 연속된 과제 전반에서 모델의 편향을 효과적으로 줄이고 안정적인 성능을 유지했으며, 이를 통해 간단하지만 효과적인 베이스라인으로 활용될 수 있음을 확인하였다.

본 논문은 이와 같은 연구를 통해 심층 신경망 기반 분류 문제에서 허위 상관으로 인한 공정성, 강건성, 편향 전이에 대한 이해를 넓히고, 다양한 분야에 적용 가능한 실용적인 해결책을 제시한다. 이는 향후 공정하고 강건하며 적응력 있는 기계학습 시스템 개발을 위한 중요한 기반이 될 것이다.

목차 (Table of Contents)

• Chapter 1 Introduction 1
• 1. Deep Neural Networks under Spurious Correlations: A Growing Concern 1
• 2. Core Challenges: Group Fairness, Group Robustness, and Bias Transfer 6
• 1) Group Fairness 6
• 2) Group Robustness 10
• 3) Bias Transfer 13
• 3. Organization of the Dissertation 15
• 1) Group Fairness in Natural Image Classification 16
• 2) Group Robustness in Medical Diagnosis 16
• 3) Bias Transfer in Continual Learning 17
• Chapter 2 Group Fairness in Natural Image Classification 18
• 1. Introduction 18
• 2. Related Works 22
• 3. Fairness Criterion 23
• 4. Method 24
• 1) MMD-based Regularization for MFD 26
• 2) Objective Function 28
• 3) Analysis 29
• 5. Experiments 31
• 1) Synthetic Dataset 32
• 2) Real-world Datasets 38
• 3) Ablation Study 41
• 6. Conclusion 42
• Chapter 3 Group Robustness in Brain Image Classification 44
• 1. Introduction 44
• 2. Related Works 48
• 1) Group Robustness of a Classifier 48
• 2) Early Diagnosis of MCI to AD Conversion 49
• 3) Group Robust Classification in Medical Imaging 50
• 3. Dataset Setup and Preliminary Study 51
• 1) Data Collection 51
• 2) Data Preprocessing 53
• 3) Analysis on Group Robustness of MCI to AD Conversion Prediction 54
• 4. Method 56
• 1) Adaptive Linear Modulation with Cognitive and Age Indicators 57
• 2) Age-based Decoupled Classifier 60
• 5. Experiments 61
• 1) Baseline Methods 61
• 2) Evaluation Metrics 62
• 3) Main Results 63
• 4) Ablation Study 65
• 5) Effect of the Tabular Features 66
• 6) Effectiveness of Linear Modulation 68
• 7) Architecture Variation Analysis 69
• 8) External Validation 70
• 6. Conclusion 71
• Chapter 4 Bias Transfer in Continual Learning 73
• 1. Introduction 73
• 2. Related Works 76
• 1) Continual Learning 76
• 2) Spurious Correlations and Debiased Learning 77
• 3) CL Considering Spurious Correlations 77
• 3. Experimental Setup 78
• 1) Problem Setting and Definition 78
• 2) Benchmark Datasets 80
• 3) Continual Learning and Debiasing Baselines 81
• 4. Case for CL with Two Tasks 82
• 1) Forward Transfer of the Bias 83
• 2) Backward Transfer of the Bias 84
• 3) Feature Representation Analysis 84
• 5. Case for CL with a Longer Sequence of Tasks 85
• 1) Bias Transfer in Longer Sequences 86
• 2) Accumulation of the Same Type of Dataset Bias 87
• 3) Accumulation of the Different Types of Dataset Bias 87
• 6. Bias-aware Continual Learning 89
• 1) Experimental Setup for Developing Bias-aware CL Method 89
• 2) Group-class Balanced Greedy Sampling 90
• 3) Performance Comparison 91
• 7. Conclusion 92
• Chapter 5 Conclusions and Future Directions 94
• References 98
• Appendix 126
• 1. Chapter 2: Implementation Details 126
• 1.1 Network Architecture for CIFAR-10S 126
• 1.2 Hyperparameters for Main Results 126
• 1.3 Details on AD+FitNet 127
• 1.4 Hyperparameters for t-SNE 128
• 2. Chapter 2: Results with Standard Deviations (SD) 128
• 3. Chapter 3: Implementation Details 130
• 4. Chapter 3: Per-Test Set Groupwise Accuracy from Preliminary Analysis 131
• 5. Chapter 3: Results of Prognosis Task 132
• 6. Chapter 3: Additional Results of Table 3-4 and Table 3-8 134
• 7. Chapter 3: Quality Control Procedure for NACC 135
• 8. Chapter 3: Groupwise Feature Information of NACC 135
• 9. Chapter 3: Details of Model Architectures 136
• 10. Chapter 3: LOFO Importance of DEAL for GBA 136
• 11. Chapter 4: BGS Algorithm 137
• 12. Chapter 4: Implementation Details 137
• 12.1 Model Architectures and Optimization 137
• 12.2 Implementations of Continual Learning Methods and BGS 138
• 12.3 Hyperparameter Selection for Each Result 140
• 12.4 Datasets 140
• 13. Chapter 4: Formal Definition of Difference of Classwise Accuracy (DCA) 143
• 14. Chapter 4: On the Issue of Local Spurious Features (LSF) in Domain-IL 144
• 15. Chapter 4: Additional Results 146
• 15.1 Two Task Scenarios with Different Bias Levels and Noise Type 146
• 15.2 Accuracy for CL with Two Tasks 148
• 15.3 Feature Representation Analysis for Backward Transfer of Bias 150
• 15.4 Results for Accuracy with a Longer Sequence of Tasks 150
• 15.5 Accumulation of the Same Type of Dataset Bias through Backward Transfer 151
• 15.6 Accumulation of the Different Types of Dataset Bias through Backward Transfer 152
• 15.7 Performance Comparison with Additional Baselines and Setting 155
• 16. Chapter 4: Analysis for Bias Transfer with More Realistic CL Scenarios 156