TLDR: A new deep learning framework, DAMSDAN (Distribution-Aware Multi-Source Domain Adaptation Network), has been developed to improve EEG-based emotion recognition. It addresses challenges posed by individual physiological variability and data distribution shifts across subjects and sessions. DAMSDAN uses a multi-module approach, including feature encoding, marginal distribution alignment with dynamic source weighting, and conditional distribution alignment with dual pseudo-labeling, to learn domain-invariant and class-discriminative emotion representations. Experiments on SEED, SEED-IV, and FACED datasets show that DAMSDAN consistently outperforms existing methods, achieving high accuracy in both cross-subject and cross-session scenarios.
Understanding human emotions through brain signals, specifically electroencephalography (EEG), holds immense potential for various applications, from brain-computer interfaces to mental health monitoring. However, a significant hurdle in this field is the vast physiological differences between individuals and across different recording sessions. These variations make it challenging for emotion recognition models to generalize effectively when applied to new subjects or new data from the same subject.
Researchers have introduced a novel deep learning framework called the Distribution-Aware Multi-Source Domain Adaptation Network (DAMSDAN) to tackle these challenges. This innovative approach aims to improve the accuracy and reliability of EEG-based emotion recognition in real-world scenarios where data often comes from diverse sources and conditions.
Addressing Core Challenges
DAMSDAN focuses on two primary issues in multi-source domain adaptation: First, how to effectively model the unique characteristics of data distributions from different source domains and determine their relevance to the target domain. This is crucial to prevent “negative transfer,” where irrelevant source data can actually harm the model’s performance. Second, the framework seeks to achieve a precise, category-level consistency between source and target domains, which is essential for the model to accurately distinguish between different emotions.
How DAMSDAN Works
The DAMSDAN framework is built upon three main modules that work together to achieve robust emotion recognition:
The first is the Feature Encoding (FE) Module. This module is responsible for extracting meaningful and transferable emotion-related patterns from the raw EEG signals. It uses a combination of common and domain-specific encoders to capture both shared and unique characteristics across different subjects or sessions.
Next, the Marginal Distribution Alignment (MDA) Module focuses on aligning the overall feature distributions between the source and target domains. It incorporates several strategies:
- Prototype Consistency Constraint (PCC): This mechanism helps organize the feature space by ensuring that features belonging to the same emotion category are clustered closely together, improving class separability.
- Adversarial Domain Alignment (ADA): Through a process similar to a game, a “domain discriminator” tries to tell if a feature comes from a source or target domain, while the feature encoder tries to create features that are indistinguishable. This makes the features “domain-invariant.”
- Domain-Aware Source Weighting (DASW): This is a smart strategy that dynamically assesses how similar each source domain is to the target domain. It then assigns higher weights to more relevant sources and reduces the influence of less relevant or noisy ones, effectively mitigating negative transfer.
Finally, the Conditional Distribution Alignment (CDA) Module ensures that the alignment happens at a more granular, class-specific level. It includes:
- Dual Pseudo-Label Collaboration (DPLC): For the unlabeled target data, this strategy generates highly reliable “pseudo-labels” by combining predictions from the classifier and insights from a clustering algorithm. It selects only the most confident labels, reducing errors.
- Prototype-Guided Conditional Alignment (PGCA): Using these reliable pseudo-labels and the class prototypes from the source domains, this mechanism aligns the class-level features across domains, ensuring that, for example, “happiness” in one subject’s data looks similar to “happiness” in another’s, even if their overall brain signals differ.
Impressive Results
The effectiveness of DAMSDAN was rigorously tested on three widely used EEG emotion datasets: SEED, SEED-IV, and FACED. The experiments covered both cross-subject scenarios (where the model is trained on some individuals and tested on entirely new ones) and cross-session scenarios (where the model is trained on earlier recordings and tested on later ones from the same individuals).
DAMSDAN consistently achieved leading performance across all datasets and scenarios. For instance, it reached average accuracies of 94.86% and 79.78% in cross-subject settings on SEED and SEED-IV, respectively, and 95.12% and 83.15% in cross-session settings. On the large-scale FACED dataset, it achieved an impressive 82.88% accuracy in cross-subject settings. These results demonstrate DAMSDAN’s superior generalization capability and robustness compared to existing state-of-the-art methods.
The research paper, available here, provides a detailed account of the methodology and experimental findings. The authors also conducted extensive ablation studies, which confirmed that each component of DAMSDAN plays a crucial role in its overall success.
Also Read:
- Cortical-SSM: A New Deep State Space Model for Enhanced Brain Signal Decoding
- Large AI Models Reshape Neuroscience Research and Clinical Practice
Conclusion
DAMSDAN represents a significant advancement in EEG-based emotion recognition. By intelligently addressing the challenges of individual variability and distributional shifts, it offers a robust and scalable solution for decoding nuanced emotional states from brain signals. This framework lays a strong foundation for future developments in affective brain-computer interface systems and other applications requiring accurate and adaptable emotion recognition.
