An Investigation into Learning Methods for Spiking Neural Networks in the Context of Autonomous Navigation

人工智能存在于我们的日常生活中，并覆盖全球。机器学习中的一种常见方法是使用人工神经网络，与监督学习一起使用。这些都在分类任务中实现了最先进的性能。这些网络使用模拟神经元模型，与更精确的生物物理神经元模型相比，该模型简化了动力学。这些更精确的模型被称为尖峰神经元模型，而这些模型的网络被称为尖峰神经网络(SNN)。SNN领域是当前研究的一个领域，在神经元动力学、网络结构和学习规则方面有许多悬而未决的问题，以最大限度地提高这些更接近生物现实的网络模型的性能。这个项目将考虑在基准问题中训练尖峰神经网络的不同方法，重点是自主导航。该项目的目的是探索监督、非监督和半监督学习在SNN中的使用。目标是：-回顾当前建模SNN动态和定义网络体系结构的方法。-确定要应用的候选学习算法，重点放在SNN中的非监督和半监督方法上。-在自主导航中定义基准问题，并考虑SNN控制系统的最合适的输入和输出映射。-计划、进行和分析一系列实验的数据，探索所选SNN学习算法和输入-输出映射的行为。-在博士论文中报告结果。研究方法将采用生物启发的方法，将信息从神经科学映射到基于软件的实现中，可应用于现实世界的问题。该项目将提供计算神经科学、机器学习、软件开发和实验数据分析方面的培训。软件将使用第三方软件和定制开发的软件(很可能是在Python和/或MatLab中)来构建。这项工作将有助于开发新的人工智能技术，以补充现有的人工神经网络方法。该项目与EPSRC的研究主题相一致：信通技术和人工智能技术。

Artificial intelligence is present throughout our everyday lives and has a global reach. A common approach in machine learning is the use of artificial neural networks, used with supervised learning. These have achieved state-of-the-art performance in classification tasks. These networks use analogue neuron models which have simplified dynamics compared to more biophysically accurate neuron models. These more accurate models are known as spiking neuron models, and networks of these are referred to as spiking neural networks (SNNs). The field of SNNs is an area of current research interest with a number of open questions regarding neuronal dynamics, network architecture and learning rules to maximise the performance of these more biologically realistic network models. This project will consider different approach to training spiking neural networks in benchmark problems, with a focus on autonomous navigation. The aim of the project is to explore the use of supervised, unsupervised and semi-supervised learning in SNNs. The objectives are to:-Review current approaches to modelling SNN dynamics and defining network architecture. -Identify candidate learning algorithms to apply with a focus on unsupervised and semi-supervised approaches in SNNs.-To define a benchmark problem in autonomous navigation with consideration of the most appropriate input and output mappings for an SNN control system.-To plan, undertake and analyse data from a series of experiments exploring the behaviour of the chosen SNN learning algorithms and input-output mappings.-To report the results in a PhD thesis.The research methodology will take a bio-inspired approach that maps information from neuroscience into software based implementations that can be applied to real-world problems. The project will provide training in computational neuroscience, machine learning, software development and experimental data analysis. Software will be constructed using both third party software and custom developed software (likely to be in either Python and/or MATLAB). The work will contribute to the development of new AI technologies to complement existing artificial neural network approaches. The project is aligned with EPSRC research themes: ICT and Artificial Intelligence Technologies.