[NOTE] Chapters 8 & 9 are embargoed till 18 July 2025

This thesis presents novel Machine Learning (ML) and pattern recognition algorithms based on Spiking Neural Networks (SNNs). SNNs mimic biological information processing more closely than the contemporary rate-coded neural networks (ANNs) with a premise of greater efficiency. However, since information in SNNs is represented by discrete spikes, developing robust learning algorithms is a challenge due to its non-differentiable nature. This complexity has led to overgrown structures and computationally costly learning strategies to achieve higher ML performances which negatively impacts the efficiency premise of SNNs. This research attempts to address this challenge by drawing inspiration from plasticity-based learning in the biological brain.

The brain learns complex patterns with minimal energy expenditure using various plasticity techniques. Plasticity in the brain occurs at synaptic, neuron, circuit, and region-wise levels across millisecond to annual time scales. Inspired by these plasticity functions, this thesis presents developments of adaptive SNN algorithms for spatiotemporal data modelling under batch mode and online learning. The novel algorithms proposed are tested using electroencephalogram (EEG) data in classification tasks related to mental stress, emotions, and motor movements.

The first contribution of this thesis is a basic three-layered SNN algorithm developed using Spike Time Dependent Plasticity (STDP) and an evolving classifier that learns from local and global spiking activity in a single pass. Structural plasticity (SP) was incorporated into this SNN via Differential Evolution (DE) algorithm, which optimised the hidden layer neuron population operating with STDP. The resulting architectures performed better than overgrown and undergrown structures and produced heuristics for neuron population selection. Using the network heuristics of this study, a more brain-like learning method with STDP and Intrinsic Plasticity (IP) is introduced as the second contribution of this study. IP had found to be regulating the neurons’ excitability to maintain network spiking stability in STDP setups. However, the particulars in guiding STDP+IP learning lack clarity. In this work, STDP+IP learning is guided using entropy and neuron redundancy measures. Selected neurons are pruned at the end of the training according to the use it or lose it strategy, a phenomenon also observed in the brain. In some cases, pruning improved the generalisation capability of the SNN. By developing this method further, an online learning method named Online-Neuroplasticity Spiking Neural Network (O-NSNN) was introduced as the third contribution. The O-NSNN is tested in classifying acute stress of individuals, and network behaviour is examined to find links between acute and perceived stress (PS). The highest classification accuracy was recorded at 93.63% and the lowest at 85.29%, outperforming SNN without SP. According to the knowledge extractions, high PS individuals had less sensitivity to daily acute stressors.

Experiments of this thesis aim to develop biologically plausible learning algorithms to enhance ML performance while improving efficiency. These developments are discussed at a fundamental level of network spiking behaviour. Therefore, the contributions reduce the knowledge gap between plasticity, spiking behaviour, and ML performance while providing insights into how the brain learns.