/Survey

This is a repository contains materials for future survey submission

Interpretable and Robust AI in EEG Systems: A Survey

This is a repository maintain an ever-evolving list of papers on the interpretability and robustness of AI in EEG systems, with frequent updates to ensure its comprehensiveness and timeliness.

The contribution of this survey

  1. This is the first comprehensive survey focusing on the interpretability and robustness of AI in EEG systems.
  2. We propose a novel taxonomy of interpretability and robustness for EEG systems.
  3. We summarize and highlight the emerging and most representative interpretable and robust AI works related to EEG systems.
  4. We discuss some open problems and promising directions for future EEG systems.
  5. A selected top tier paper list focusing interpretability and robustness of AI in EEG systems, and the list goes here: Paper List

Summary of Interpretable AI in EEG Systems

Interpretability Categories Methods Coverage Explanation Type
Backpropagation-based Methods LRP Local/Global Attribution
DeepLIFT Local/Global Attribution
CAM Local Attribution
Grad-CAM Local Attribution
Perturbation-based Methods LIME Local Attribution
SHAP Local Attribution
Rule-based Methods Random Forest Global Decision Rules
Fuzzy Inference Systems Global Fuzzy Rules
Bayesian Systems Global Bayesian Rules

Comparison of different interpretanility methods in EEG Systems

Backpropagation-based Method Perturbation-based Method Rule-based Method
Mechanism Analyze the feature contribution by backpropagating the gradients from predictions. Explain the original model's behavior with local surrogate models. Explain model using specific logic rules
Explanation Stage Post-hoc Post-hoc Ante-hoc
Model Dependence Model-specific Model-agnostic Model-agnostic
Flexibility Low High High
Application Scenario Differentiable models Tolerable of high computational costs Availability of priori knowledge
Limitation Gradient dependency; Narrow applicability Computationally intensive; Prone to overfitting Oversimplify complex EEG systems; Require domain expertise

Summary of Robust AI in EEG Systems

Undesirable Factors Subcategory Methods
Noise and Artifacts External Noise Traditional Signal Processing
Internal Artifacts Models' Self-Robustness
Human Variability Cross-subject Issues Transfer Learning , Dynamic Domain Adaptation
Cross-session Issues Transfer Learning , Robust Feature Extraction
Data Acquisition Instability Resistance Change Attention Mechanism
Channel Missing & Broken Missing Data Reconstruction
Adversarial Attacks Evasion & Manipulation Adversarial Training