A Spiking Neural Network Based Pipeline for Motor Imagery EEG Classification

Abstract

Motor imagery based electroencephalogram signals have been used for developing brain-computer interfaces because of their ease of execution and low-risk profile. These systems have achieved reliable performances with the emergence of machine learning. However, these machine-learning approaches are not focused on optimizing energy efficiency. Yet, energy efficiency is crucial for implementing portable devices since the continuous operational time depends on it. The spiking neural network based classifiers are suitable and promising for achieving energy efficiency compared to the conventional machine learning algorithms while maintaining the same level of accuracy. Unlike traditional systems, where all computing units operate at each clock cycle, spike based systems process information asynchronously only when there are sufficient inputs to activate neurons; thus, these algorithms can be utilized on specialized hardware chips which can process spikes for achieving higher energy efficiency. The primary objective of this thesis was to develop a spiking neural network based pipeline that incorporates a spatial filtering method called the common spatial pattern. Since motor imagery tasks are highly correlated with unique spatial areas, the common spatial pattern method can improve the separability between the motor imagery classes by enhancing the spatial features of the electroencephalogram channels. With a higher separability, input signals can be effectively encoded into spikes without losing the critical information for classification tasks. To achieve this objective, three types of pipelines (rate coding, latency coding and delta coding) were implemented to classify the labelled (“left hand” and “right hand”) motor imagery signals in a widely used open electroencephalogram dataset. In addition, to evaluate the contribution of the common spatial pattern method comparatively, another set of pipelines was implemented with the same steps, excluding the use of the common spatial pattern method. By using each pipeline, the classifier models were trained on different datasets (five individual and one multi-subject dataset) using a surrogate gradient decent based back propagation algorithm. 5-fold cross-validation was used to validate the performance of each pipeline model on each dataset. Each trained model has demonstrated acceptable classification accuracy with the common spatial pattern method compared to conventional methods. Among them, the rate coding pipeline models recorded the highest mean accuracies on training and testing datasets, while latency coding pipeline models reflected the best energy efficiency. Results from the pipelines which do not utilize the common spatial pattern showed a drastic decrease in performance, highlighting the effectiveness of spatially filtered signals for the overall classification process. These findings direct future research to explore spike-friendly spatial filtering techniques and novel spiking neural network architectures to further enhance classification accuracy and energy efficiency in motor imagery electroencephalogram classification.