Evolving probabilistic spiking neural networks for modelling and pattern recognition of spatio-temporal data on the case study of Electroencephalography (EEG) brain data
MetadataShow full metadata
The use of Electroencephalography (EEG) in Brain Computer Interface (BCI) domain presents a challenging problem due to presence of spatial and temporal aspects inherent in the EEG data. Many studies either transform the data into a temporal or spatial problem for analysis. This approach results in loss of significant information since these methods fail to consider the correlation present within the spatial and temporal aspect of the EEG data. However, Spiking Neural Network (SNN) naturally takes into consideration the correlation present within the spatiotemporal data. Hence by applying the proposed SNN based novel methods on EEG, the thesis provide improved analytic on EEG data. This thesis introduces novel methods and architectures for spatio-temporal data modelling and classification using SNN. More specifically, SNN is used for analysis and classification of spatiotemporal EEG data. In this thesis, for the first time, Ben Spiker Encoder Algorithm (BSA) is applied on EEG data and its applicability is demonstrated successfully. Moreover, three new stochastic neural models are introduced; namely, Stochastic Noisy Reset (NR), Stochastic step-wise noisy threshold model (ST) and Continuous stochastic threshold (CT). The stochastic models mimic activity of biological neurons while retaining low implementation cost. Also, a modification of precise-time spike pattern association neuron (SPAN) called stochastic precise-time spike pattern association neuron (SSPAN) is introduced. The SSPAN demonstrates superior performance than SPAN, especially when used with stochastic neural models. A novel Dynamic Evolving Probabilistic Spiking Neural Network (DepSNN) is introduced as an extension of the eSNN model. Five novel variants of DepSNN (DepSNNm, DepSNNs, NR-DepSNNs, ST-DepSNNs, and CT-DepSNNs) are presented and the results show that it requires high density of input spikes if SDSP learning is to be made efficient. The thesis then offers a critical analysis of Electroencephalography (EEG) data analysis and classification methods used to date. The developed SNN methods have been adopted in EEG analysis and classification investigated on two datasets (real-world audio-visual stimuli perception EEG dataset and P300 based BCI dataset), with promising results relative to other methods. Furthermore, the proposed novel SNN architecture for spatio-temporal data termed evolving probabilistic SNN reservoir (epSNNr) shows enhanced performance when integrated with stochastic neural models. The utilization of 3D Localization mapping along with DepSNN as a readout unit, showed very outstanding results especially on P300 based BCI application.