Thesis

Once upon a spike time: exploring spiking neurons and perspective applications of neuromorphic solutions

Downloadable Content

Download PDF
Creator
Rights statement
Awarding institution
  • University of Strathclyde
Date of award
  • 2025
Thesis identifier
  • T17230
Person Identifier (Local)
  • 202087149
Qualification Level
Qualification Name
Department, School or Faculty
Abstract
  • Despite their outstanding achievements in a number of different fields, conventional neural network-based solutions often fall short in terms of concrete applicability to edge devices and systems with limited autonomy or computational resources. This is due to the intensive computations that are required to process input data and the size of modern neural networks, which also test the memory requirement of a given setup. Neuromorphic (NM) computing offers a shift in this paradigm by enabling low-power Artificial Intelligence (AI) through sparse and localized computations, and low latency thanks to its asynchronicity. Spiking Neural Networks (SNN), with spiking neurons at their core, represent the algorithmic foundation of Neuromorphic AI. In recent years, the research on SNNs has been highly proactive with several advancements being proposed in various contexts, in an effort to try and bridge the performance gap with conventional approaches that has traditionally characterized the field of NM computing. Whilst research efforts are showing promising results, there is still a lack of shared foundational knowledge and understanding of the interplay between different components, particularly the spiking neuron models. Furthermore, a reason for debate is given by what represents a good task for the SNNs to be evaluated given that more traditional ones might not fully highlight the representational abilities of SNNs. As such, new directions of research are always sought, where new tasks that could benefit from the use of SNNs’ feature extraction abilities are explored. This thesis aims to address the above points as a means to help advance the overall knowledge in the field of Neuromorphic computing with SNNs. The starting point of the investigations resides in the understanding of the importance of spiking neurons in an NM machine learning (ML) pipeline. In the neuroscience literature, spiking neuron models are extensively discussed as they are the means by which scientists model the human brain. Such mathematical models determine how incoming information affects the internal state of a neuron in relation to one or more variables, and regulate the emission of spikes, the signals that neurons use to communicate information. The same level of attention is not normally paid when employing neuron models in engineering pipelines and, often, the simplest model, the Leaky Integrate-and-Fire (LIF), is used for its simplicity and efficiency. Starting from a selection of three relatively simple models, this thesis presents a study aimed at highlighting whether the choice of any neuron model is more appropriate with respect to another one. This is tested within a simple framework where an SNN is trained with Spike Time-Dependent Plasticity (STDP), a biologically inspired unsupervised learning rule. The tasks presented to the SNN are image classification tasks based on NM datasets of increasing difficulty. The study reveals that higher levels of complexity in the neuronal dynamics can in fact prove beneficial where the complexity of the temporal features in the data is also higher. The thesis continues with an exploration of a possible approach to a field that is not often considered for NM applications: that of time series forecasting. The proposed approach encompasses two main aspects: a novel neuron population encoding system, and two novel bespoke loss functions. The encoding system leverages concepts from the differencing transform used in time series analysis and gets inspiration from neuromorphic vision sensors. In this way, the encoded signal is not only rendered more stationary and amenable to processing but it is also shown to approximate the derivative of the signal itself. The proposed loss functions build upon biological concepts and from the knowledge about the encoding step. The overall solution comprising the encoding system and one of the proposed loss functions is shown to outperform the reference system, thereby demonstrating the potential of SNNs to be applied in this domain. Finally, a novel solution to surface Electromyography (sEMG) gesture classification is presented. sEMG is a crucial technology frequently utilized in health-related applications, including prosthetics, where the advantages brought by NM engineering could be highly beneficial. The solution comprises the use of Resonate-and-Fire (RF) neurons, a type of neuron that has been shown to approximate a Short Time Fourier Transform (STFT), to encode EMG signals into a spike-frequency domain whilst performing filtering on the unwanted frequencies. This is paired with a novel decoding approach that leverages a convolutional layer to transform the signal back into the temporal domain, thereby enabling hyperparameter optimization on the encoding neurons. The overall solution is tested on a challenging dataset and is shown to outperform reference works on the same task. This final piece of research not only underscores the importance of the selection of appropriate spiking neurons for a given task but also how this can enable closing the performance gap between NM computing and conventional methods whilst maintaining the advantages of neuromorphic technologies.
Advisor / supervisor
  • Di Caterina, Gaetano
Resource Type
DOI

Relations

Items

Thumbnail Title Date Uploaded Visibility Actions
file details PDF of thesis T17230 2025-04-11 Public Download