A novel data-free class incremental learning framework with adaptive sparse reasoning for EEG emotion recognition

In the framework of incremental learning based on electroencephalography (EEG), cross-subject samples of the same emotion category tend to exhibit highly similar feature distributions. Such homogeneity frequently triggers catastrophic forgetting during model updates. While replaying historical samples or features may alleviate this issue, sample replay requires model retraining, and feature replay typically extracts only terminal-layer features containing limited discriminative information for classification. Therefore, we propose a data-free class incremental learning framework with adaptive sparse reasoning (DCIL-ASR). Specifically, a feature extraction network incorporating Alterable Kernel Convolution (AKConv) and Adaptive Sparse Feature Refinement (ASFR) module is designed to accommodate incremental learning tasks characterized by continuous sample updates. AKConv dynamically constructs diverse sampling structures to enrich feature extraction patterns, while ASFR enhances sparse attention mechanisms to focus on core discriminative features. Subsequently, embedded distillation is employed to extract discriminative features concealed within the intermediate layers of the network, enabling the classification knowledge relevant to the current task to be fully captured. Additionally, the feature generation method of the variational autoencoder is optimized using discriminative knowledge, guiding the model to effectively revisit a substantial set of features that encapsulate task-related discrimination knowledge. By relying on a rich repository of discriminative knowledge and a task-oriented feature generation strategy, the network is endowed with sufficient learning and memory capabilities, thereby eliminating its dependence on original samples. Under the condition of no saved samples, the DCIL-ASR achieved an emotion recognition accuracy of 83.5% for 22 categories and 81.4% for 23 categories, demonstrating its effectiveness in incremental learning.