MAMBA STATE SPACE MODEL: A COMPREHENSIVE EVALUATION OF ITS EFFICACY IN BRAIN-COMPUTER INTERFACE USING EEG

Abstract

The increasing complexity of EEG data analysis in Brain-Computer Interface (BCI) systems calls for robust and scalable machine learning models capable of extracting meaningful spatiotemporal patterns. This project evaluates Mamba SSM, a novel deep learning architecture rooted in State Space Models (SSMs), for EEG signal classification. Building on the limitations of Transformers Mamba incorporates bidirectional Mamba blocks to address challenges such as computational overhead, long-sequence processing, and noise in EEG signals. The project includes an extensive literature review comparing Mamba and Transformer based architectures for EEG analysis, highlighting their strengths and weaknesses. Experimental development involved adapting Mamba for EEG datasets, focusing on architectural enhancements, and hardware efficiency. Initial results demonstrate the model’s competitive accuracy in motor imagery classification tasks and its potential for scalability in real-time BCI applications. Challenges such as EEG data variability, dependency management, and computational resource limitations informed iterative development and learning outcomes. These included a deeper understanding of EEG datasets’ structures, architecture design for time-series data, model training on resource-constrained systems and dependencies settings. Discussions center on Mamba’s balance of efficiency and accuracy, emphasizing its utility for edge-device deployment in assistive technologies. This study investigates the use of Mamba State Space Model in the Domain of BCI Motor Imagery classification tasks. During the investigation, several architectures were proposed either based on pure Mamba blocks or combined with Attention layers. The findings suggest that Mamba SSM holds promise for advancing BCI systems, with future work exploring domain-specific optimizations, self-supervised learning, and broader applications in neuroscience.