An Incremental Learning Framework for Dynamic Neural Spike Classification Based on Attribute Self-Interaction Networks
Renzhu Wang, Pengbo Zhao, Yu Ma · 2025
Dynamic neural spike classification is an important task in decoding neural activity in the brain. It requires a system that can process spike signals from neurons in real time and accurately classify new categories as they appear. However, current classification systems are usually difficult to handle class-imbalanced data especially when the number of classes is changing. To address this problem, an incremental learning framework for dynamic neural spike classification based on attribute self-interaction networks is proposed. Through the incremental learning of dynamic output expansion, self-interaction feature augmentation, and open set detection, this framework effectively copes with the problems of category increasing and catastrophic forgetting. Specifically, it mitigates the forgetting problem by using reservoir sampling techniques to store samples from old tasks and replaying them during training. Experimental results show that the framework performs better compared to state-of-the-art methods, especially when the number of classes is changing. Thus, it is important to real-world applications, such as continuous monitoring or automated diagnosis for neurological diseases.