腦電圖（EEG, Electroencephalography）是一種有效捕捉大腦活動的工具，但跨受試者、任務和資料集的變異性為建立可泛化模型帶來了重大挑戰。傳統的機器學習方法往往無法應對這種變異性，導致在新情境中的適應能力較差。為了解決這個問題，我們提出了一個創新的元學習框架，將混合專家（MoE, Mixture of Expert）架構與跨資料集任務抽樣策略結合，以提高腦電圖分類的泛化能力。
我們的方法整合了多個專家網絡，每個專家專注於不同的任務類型，並且有一個門控(gating)網絡根據支持集(support set)動態選擇專家。通過從多個資料集中取樣任務，模型被鼓勵學習與任務無關的表徵，並能適應多樣化的腦電圖任務。這個訓練框架模擬了少樣本學習，並允許在元測試期間對新任務進行快速適應。
我們在一系列基於事件誘發電位（ERP, Event-related potential）和滑動視窗腦電圖任務上評估了這個方法。結果顯示，所提出的方法在多個事件誘發電位的任務中優於基線，顯示出在事件相關電位任務的泛化能力有不錯的結果。然而，在更複雜的情境中，性能較不穩定，例如在Fingerspelling任務或滑動視窗腦電圖任務中，較相似的類別以及連續腦電圖信號對門控機制帶來了困難。而在消融研究中強調了混合專家架構的設計和跨資料集任務抽樣的重要性，但同時也揭示了當前門控機制在複雜情境下的限制。
而除了架構上的創新外，本研究還凸顯了一項設計：專家模型的效能受到們控機制的限制。因此，未來應著重開發更精細且具感知時間能力的門控機制，例如採用長短期記憶模型(Long Short-Term Memory, LSTM)或注意力機制(attention mechansims)來克服平均池化(mean-pooling)的限制。此外，結合心跳誘發電位(heartbeat-evoked potentials)，亦是一個增加穩定性的潛在方向。整體而言，本研究不僅提供了一套新穎的框架，也為跨異質資料集的腦電圖分類提供了重要的啟發與方向。

Electroencephalography (EEG) is a valuable tool for capturing brain activity, yet the variability across subjects, tasks, and datasets poses a major challenge for building generalizable models. Traditional machine learning methods often struggle with such variability, leading to poor adaptability in new contexts. To address this, we propose a novel meta-learning framework that combines a Mixture of Experts (MoE) architecture with a cross-dataset task sampling strategy, aiming to improve generalization in EEG classification.
Our method integrates multiple expert networks, each specializing in different task types, and a gating network that dynamically selects experts based on the support set. By sampling tasks from multiple datasets, the model is encouraged to learn task-agnostic representations and adapt to diverse EEG tasks. The episodic training framework simulates few-shot learning and allows rapid adaptation to new tasks during meta-testing.
We evaluate the method across a range of Event-Related Potential (ERP)-based and windowed continuous EEG tasks. The results show that the proposed method outperforms baseline models in several ERP-based tasks, demonstrating enhanced generalization in structured event-related paradigms. However, performance is less consistent in more complex settings, such as the Fingerspelling task or windowed continuous EEG, where subtle label distinctions and the inherent challenges of continuous EEG signals pose difficulties for the gating mechanism. Ablation studies underscore the significance of the MoE design and cross-dataset sampling strategy, while also revealing limitations of the current gating mechanism under complex conditions.
Beyond architectural innovations, this work highlights a key design principle: the performance of expert-based models is bounded by the information content of their routing inputs. Therefore, future directions should focus on developing more sophisticated, temporally-aware gating mechanisms—such as those incorporating Long Short-Term Memory (LSTM) networks or attention mechanisms—to overcome mean-pooling limitations. Additionally, integrating complementary modalities like ECG through heartbeat-evoked potentials presents a promising avenue to enhance physiological state characterization and improve classification robustness.
Overall, this study not only introduces a novel framework but also provides critical insights and future directions for advancing EEG classification across heterogeneous tasks and datasets.

[1] A. Delorme and S. Makeig, “Eeglab: an open source toolbox for analysis of singletrial eeg dynamics including independent component analysis,” Journal of neuroscience methods, vol. 134, no. 1, pp. 9–21, 2004.
[2] X. Gu, Z. Cao, A. Jolfaei, P. Xu, D. Wu, T.-P. Jung, and C.-T. Lin, “Eeg-based braincomputer interfaces (bcis): A survey of recent studies on signal sensing technologies and computational intelligence approaches and their applications,” IEEE/ACM transactions on computational biology and bioinformatics, vol. 18, no. 5, pp. 1645–1666, 2021.
[3] D. Wu, Y. Xu, and B.-L. Lu, “Transfer learning for eeg-based brain–computer interfaces: A review of progress made since 2016,” IEEE Transactions on Cognitive and Developmental Systems, vol. 14, no. 1, pp. 4–19, 2020.
[4] X. Li, Y. Zhang, P. Tiwari, D. Song, B. Hu, M. Yang, Z. Zhao, N. Kumar, and P. Marttinen, “Eeg based emotion recognition: A tutorial and review,” ACM Computing Surveys, vol. 55, no. 4, pp. 1–57, 2022.
[5] Y. Badr, U. Tariq, F. Al-Shargie, F. Babiloni, F. Al Mughairbi, and H. Al-Nashash, “A review on evaluating mental stress by deep learning using eeg signals,” Neural Computing and Applications, vol. 36, no. 21, pp. 12 629–12 654, 2024.
[6] R. Richer, N. Zhao, J. Amores, B. M. Eskofier, and J. A. Paradiso, “Real-time mental state recognition using a wearable eeg,” in 2018 40th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). IEEE, 2018, pp. 5495–5498.
[7] D. Brandeis and D. Lehmann, “Event-related potentials of the brain and cognitive processes: approaches and applications,” Neuropsychologia, vol. 24, no. 1, pp. 151–168, 1986.
[8] P. Montoya, R. Schandry, and A. Müller, “Heartbeat evoked potentials (hep): topography and influence of cardiac awareness and focus of attention,” Electroencephalography and Clinical Neurophysiology/Evoked Potentials Section, vol. 88, no. 3, pp. 163– 172, 1993.
[9] C. Finn, P. Abbeel, and S. Levine, “Model-agnostic meta-learning for fast adaptation of deep networks,” in International conference on machine learning. PMLR, 2017, pp. 1126–1135.
[10] J. Snell, K. Swersky, and R. Zemel, “Prototypical networks for few-shot learning,” Advances in neural information processing systems, vol. 30, 2017.
[11] O. Vinyals, C. Blundell, T. Lillicrap, D. Wierstra et al., “Matching networks for one shot learning,” Advances in neural information processing systems, vol. 29, 2016.
[12] A. Santoro, S. Bartunov, M. Botvinick, D. Wierstra, and T. Lillicrap, “Meta-learning with memory-augmented neural networks,” in International conference on machine learning. PMLR, 2016, pp. 1842–1850.
[13] T. Munkhdalai and H. Yu, “Meta networks,” in International conference on machine learning. PMLR, 2017, pp. 2554–2563.
[14] H. W. Ng and C. Guan, “Subject-independent meta-learning framework towards optimal training of eeg-based classifiers,” Neural Networks, vol. 172, p. 106108, 2024.
[15] S. Li, H. Wu, L. Ding, and D. Wu, “Meta-learning for fast and privacy-preserving source knowledge transfer of eeg-based bcis,” IEEE Computational Intelligence Magazine, vol. 17, no. 4, pp. 16–26, 2022.
[16] M. Liu, J. Liu, M. Xu, Y. Liu, J. Li, W. Nie, and Q. Yuan, “Combining meta and ensemble learning to classify eeg for seizure detection,” Scientific Reports, vol. 15, no. 1, p. 10755, 2025.
[17] C. Chen, H. Fang, Y. Yang, and Y. Zhou, “Model-agnostic meta-learning for eeg-based inter-subject emotion recognition,” Journal of Neural Engineering, vol. 22, no. 1, p. 016008, 2025.
[18] J.-W. Han, S. Bak, J.-M. Kim, W. Choi, D.-H. Shin, Y.-H. Son, and T.-E. Kam, “Metaeeg: Meta-learning-based class-relevant eeg representation learning for zero-calibration brain–computer interfaces,” Expert Systems with Applications, vol. 238, p. 121986, 2024.
[19] R. A. Jacobs, M. I. Jordan, S. J. Nowlan, and G. E. Hinton, “Adaptive mixtures of local experts,” Neural computation, vol. 3, no. 1, pp. 79–87, 1991.
[20] N. Shazeer, A. Mirhoseini, K. Maziarz, A. Davis, Q. Le, G. Hinton, and J. Dean, “Outrageously large neural networks: The sparsely-gated mixture-of-experts layer,” arXiv preprint arXiv:1701.06538, 2017.
[21] T. Hospedales, A. Antoniou, P. Micaelli, and A. Storkey, “Meta-learning in neural networks: A survey,” IEEE transactions on pattern analysis and machine intelligence, vol. 44, no. 9, pp. 5149–5169, 2021.
[22] A. Nichol, J. Achiam, and J. Schulman, “On first-order meta-learning algorithms,” arXiv preprint arXiv:1803.02999, 2018.
[23] S. Koelstra, C. Muhl, M. Soleymani, J.-S. Lee, A. Yazdani, T. Ebrahimi, T. Pun, A. Nijholt, and I. Patras, “Deap: A database for emotion analysis; using physiological signals,” IEEE transactions on affective computing, vol. 3, no. 1, pp. 18–31, 2011.
[24] S. Katsigiannis and N. Ramzan, “Dreamer: A database for emotion recognition through eeg and ecg signals from wireless low-cost off-the-shelf devices,” IEEE journal of biomedical and health informatics, vol. 22, no. 1, pp. 98–107, 2017.
[25] W.-L. Zheng and B.-L. Lu, “Investigating critical frequency bands and channels for eeg-based emotion recognition with deep neural networks,” IEEE Transactions on autonomous mental development, vol. 7, no. 3, pp. 162–175, 2015.
[26] R.-N. Duan, J.-Y. Zhu, and B.-L. Lu, “Differential entropy feature for eeg-based emotion classification,” in 2013 6th international IEEE/EMBS conference on neural engineering (NER). IEEE, 2013, pp. 81–84.
[27] T. Haarnoja, A. Zhou, P. Abbeel, and S. Levine, “Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor,” in International conference on machine learning. Pmlr, 2018, pp. 1861–1870.
[28] K. Sechidis and G. Brown, “Simple strategies for semi-supervised feature selection,” Machine Learning, vol. 107, no. 2, pp. 357–395, 2018.
[29] B. Lee, S. E. Ortega, P. M. Martinez, K. J. Midgley, P. J. Holcomb, and K. Emmorey, “”neural associations between fingerspelling, print, and signs: An erp priming study with deaf readers”,” 2024.
[30] J. R. Girard, A. M. Bishop, and C. D. Hassall, “”song familiarity”,” 2025.
[31] B. Blankertz, G. Dornhege, M. Krauledat, K.-R. Müller, and G. Curio, “The noninvasive berlin brain–computer interface: fast acquisition of effective performance in untrained subjects,” NeuroImage, vol. 37, no. 2, pp. 539–550, 2007.
[32] K. Ono, J. Hashimoto, R. Hiramoto, T. Sasaoka, and S. Yamawaki, “Modulatory effects of prediction accuracy on electroencephalographic brain activity during prediction,” Frontiers in Human Neuroscience, vol. 15, p. 630288, 2021.
[33] M. G. Machizawa, G. Lisi, N. Kanayama, R. Mizuochi, K. Makita, T. Sasaoka, and S. Yamawaki, “Quantification of anticipation of excitement with a three-axial model of emotion with eeg,” Journal of Neural Engineering, vol. 17, no. 3, p. 036011, 2020.
[34] I. J. Bajwa1, A. S. Nilsen1, . René Skukies1, A. Aamodt1, G. Ernst2, J. F. Storm1, and . Bjørn E. Juel1, “”a repeated awakening study exploring the capacity of complexity measures to capture dreaming during propofol sedation”,” 2024.
[35] P. J. Lang, M. M. Bradley, B. N. Cuthbert et al., International affective picture system (IAPS): Affective ratings of pictures and instruction manual. NIMH, Center for the Study of Emotion & Attention Gainesville, FL, 2005.
[36] F. Perrin, J. Pernier, O. Bertrand, and J. F. Echallier, “Spherical splines for scalp potential and current density mapping,” Electroencephalography and clinical neurophysiology, vol. 72, no. 2, pp. 184–187, 1989.
[37] Y. Tanaka, Y. Ito, M. Shibata, Y. Terasawa, and S. Umeda, “Heartbeat evoked potentials reflect interoceptive awareness during an emotional situation,” Scientific reports, vol. 15, no. 1, p. 8072, 2025.
[38] A. Gentsch, A. Sel, A. C. Marshall, and S. Schütz-Bosbach, “Affective interoceptive inference: Evidence from heart-beat evoked brain potentials,” Human Brain Mapping, vol. 40, no. 1, pp. 20–33, 2019.
[39] O. Pollatos, B. M. Herbert, S. Mai, and T. Kammer, “Changes in interoceptive processes following brain stimulation,” Philosophical Transactions of the Royal Society B: Biological Sciences, vol. 371, no. 1708, p. 20160016, 2016.
[40] V. J. Lawhern, A. J. Solon, N. R. Waytowich, S. M. Gordon, C. P. Hung, and B. J. Lance, “Eegnet: a compact convolutional neural network for eeg-based brain–computer interfaces,” Journal of neural engineering, vol. 15, no. 5, p. 056013, 2018.
[41] J. Puigcerver, C. Riquelme, B. Mustafa, and N. Houlsby, “From sparse to soft mixtures of experts,” arXiv preprint arXiv:2308.00951, 2023.
[42] B. Ermis, G. Zappella, and C. Archambeau, “Towards robust episodic meta-learning,” in Uncertainty in Artificial Intelligence. PMLR, 2021, pp. 1342–1351.
[43] H. Liu, C. Chen, and T. Zhang, “Face: few-shot adapter with cross-view fusion for cross-subject eeg emotion recognition,” arXiv preprint arXiv:2503.18998, 2025.
[44] V. Jayaram, M. Alamgir, Y. Altun, B. Scholkopf, and M. Grosse-Wentrup, “Transfer learning in brain-computer interfaces,” IEEE Computational Intelligence Magazine, vol. 11, no. 1, pp. 20–31, 2016.
[45] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural computation, vol. 9, no. 8, pp. 1735–1780, 1997.
[46] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in neural information processing systems, vol. 30, 2017.
[47] M. Gori, G. Monfardini, and F. Scarselli, “A new model for learning in graph domains,” in Proceedings. 2005 IEEE international joint conference on neural networks, 2005., vol. 2. IEEE, 2005, pp. 729–734.