Deep Learning and Knowledge Representation in Brain-Inspired Spiking Neural Networks for Brain-Computer Interfaces
Kumarasinghe, Kumara Vidanalage Dona Chithrangi Kaushalya
MetadataShow full metadata
Brain-Computer Interfaces aim at decoding neural commands from neurological signals and translate them into machine commands for manipulating digital devices. It provides a way of bypassing affected neural pathways in people with movement impairments. A growing body of literature on non-invasive Brain-Computer Interfaces for motor recovery and restoration highlights the need for improving the machine learning methods that decode neural activity from EEG signals. The low accuracy in decoding movements of the same limb, less biological plausibility, lack of interpretability, high prediction latency, low degree of freedom are some of the significant drawbacks in existing machine learning models used in restorative Brain-Computer Interfaces. This thesis proposes a Brain-Inspired Spiking Neural Network (BI-SNN) model for incremental learning of spike sequences from stochastic data streams as a promising step towards developing intelligent machines for Brain-Computer Interfaces. The proposed BI-SNN is a generic SNN architecture that can be applied for the predictive modelling of spatio-temporal data streams. Here it was applied to construct an interpretable neural decoder which can incrementally learn spike sequences from Electroencephalography signals. The thesis suggests that the proposed Spiking Neural Network approach results in a better neural decoder compared to the traditional machine learning approaches used by restorative BCIs in multiple aspects. The thesis proposes two spike-based learning algorithms that extended the generic NeuCube SNN framework to address seven research questions. A series of experiments were performed to address these research questions and to benchmark the model performance with multiple machine learning models. In the first study, the thesis demonstrates the feasibility of proposed eSPANNet learning algorithm to learn complex spike sequences from stochastic data streams. As an evolving model, the eSPANNet does not require certain predefined parameters related to network architecture, such as the number of neurons in the hidden layer, as it evolves neurons if needed. In the second study, the thesis presents a theoretical framework, algorithmic pipeline and associated software for representing and extraction of deep knowledge from Spiking Neural Networks for enhancing the interpretability of SNN. In the third study, the thesis integrates the proposed learning algorithms with the generic NeuCube SNN framework for constructing a novel Brain-Inspired Brain-Computer Interface. The thesis revealed that the integration of eSPANNet with the NeuCube SNN architecture could gain a higher accuracy than the standalone sensor-space eSPANNet architecture. The study benchmarked the performance of the proposed learning algorithms and showed a statistically significant improvement in prediction accuracy than several machine learning methods. The thesis has shown the feasibility of extracting neural information that contributes to controlling a wide range of motor parameters such as muscle activity and joint kinematics from Electroencephalography using the proposed BI-SNN in healthy people. In conclusion, this approach has shown the potential to construct an interpretable neural decoder which can incrementally learn to predict complex movements in real-time from Electroencephalography. This study is one of the first attempts to examine the feasibility of finding neural correlates of muscle activity and kinematics from Electroencephalography using a brain-inspired computational paradigm.