This thesis aims to build a deep spiking neural network (SNN) using spike-based filters for temporal feature extraction to produce distinguishing spiking activity while classifying spatiotemporal data. Existing deep learning (DL) architectures such as convolutional neural networks (CNNs) combined with long short-term memory (LSTM) neural networks have demonstrated their ability to classify this type of data. The neurons employed in these DL architectures do not include a time component in their operation modules resulting in loss of possibly important temporal information. The spiking neurons of third-generation neural networks (NNs), i.e., SNNs, produce discrete eventbased output incorporating time dimension in their computational model. However, the application of SNNs to produce temporal-based feature extraction and classification is still not well understood.

Current DL architectures, such as CNNs, use hierarchical feature representation to achieve high classification accuracy. Recent studies have adopted a similar design approach to implement variations of deep SNNs. These artificial neural networks (ANN) to SNN conversions are challenging as they use approximation techniques for producing spikes in SNNs that may result in the loss of temporal feature information. In short, it is not currently known how to build SNN architectures that combine the advantages of both CNNs and SNNs such that temporal features can be extracted layer by layer within a spiking neural framework.

To address the above-mentioned limitations, this thesis proposes a novel temporalbased SNN architecture that is shown to be effective when applied to spatiotemporal datasets, i.e., electroencephalogram (EEG) and music signals. In addition, a method for hand-engineered spike-based filters for use in deep SNNs to extract spatiotemporal information from the output of spiking neurons is proposed.

The first contribution of this thesis is the applicability of utilising the neuron’s voltage and spike timings as classification criteria. Next, spike train analysis methods reveal the value of information hidden in the temporal coding neural scheme. The final contribution of this research is building a deep SNN framework with hierarchical feature representation, using rate and temporal neural coding techniques. Furthermore, the proposed end-to-end SNN design presented in this research is optimised to be applied in clinical management.