基于MAML的少样本跨被试事件相关电位分类快速适应方法

MAML-Based Rapid Adaptation for Few-Shot Cross-Subject ERP Classification

摘要: 针对跨被试事件相关电位(ERP)解码中个体间差异大、新用户校准样本极少导致传统方法泛化性能严重下降的问题，文章提出一种基于模型无关元学习(MAML)的少样本快速适应方法DF-MAML。该方法以预训练的域不变特征提取器为初始参数，通过元学习框架在源被试构建的大量模拟任务上进行双层优化：内循环利用支持集模拟快速校准，外循环优化元参数使其具备“易于快速适应”的特性。在此基础上，进一步提出基于任务难度的动态分类器冻结策略，根据支持集与查询集的分布偏移自适应地决定是否冻结分类器梯度，从而稳定特征学习并提升元参数质量。在公开ERN数据集与自采中文语义句法违反数据集上的实验表明，在5-shot设定下，本方法的平均AUC分别达到0.7563和0.6705，显著优于EEGNet、原型网络及标准MAML等基线方法；消融实验验证了动态冻结策略的有效性；少样本性能曲线显示，提供5个校准样本可使性能较零样本提升约9%。本方法为低校准负担、高鲁棒性的脑机接口系统提供了关键技术支撑。

Abstract: To address the challenge in cross-subject event-related potential (ERP) decoding, where large inter-subject variability and extremely limited calibration samples for new users severely degrade the generalization performance of conventional methods, this paper proposes a MAML-based few-shot rapid adaptation approach. The proposed method initializes with a pre-trained domain-invariant feature extractor and performs bi-level optimization over a large number of simulated tasks constructed from source subjects within a meta-learning framework. Specifically, the inner loop leverages the support set to emulate rapid calibration, while the outer loop optimizes meta-parameters to facilitate fast adaptation. Furthermore, a task-difficulty-aware dynamic classifier freezing strategy is introduced, which adaptively decides whether to freeze classifier updates based on the distribution shift between the support and query sets, thereby stabilizing feature learning and improving the quality of the meta-learned parameters. Experiments conducted on a public ERN dataset and a self-collected Chinese semantic-syntax violation dataset demonstrate that, under the 5-shot setting, the proposed method achieves average AUC scores of 0.7563 and 0.6705, respectively, significantly outperforming baseline methods including EEGNet, Prototypical Networks, and standard MAML. Ablation studies further validate the effectiveness of the proposed dynamic freezing strategy. The few-shot performance analysis shows that using only five calibration samples yields an improvement of approximately 9% over the zero-shot setting. The proposed method offers a promising solution for developing brain-computer interface systems with low calibration cost and high robustness.