Unlocking spiking neural networks for machine learning research

点击翻译按钮获取中文摘要

In the last decade there has been an explosion in artificial intelligence research in which artificial neural networks, emulating biological brains, are used to solve problems ranging from obstacle avoidance in self-driving cars to playing complex strategy games. This has been driven by mathematical advances and powerful new computer hardware which has allowed large 'deep networks' to be trained on huge amounts of data. For example, after training a deep network on 'ImageNet' - which consists of over 14 million manually annotated images - it can accurately identify the content of images. However, while these deep networks have been shown to learn similar patterns of connections to those found in the parts of our brains responsible for early visual processing, they differ from real brains in several important ways, especially in how the individual neurons communicate. Neurons in real brains exchange information using relatively infrequent electrical pulses known as 'spikes', whereas, in typical artificial neural network models, the spikes are abstracted away and values representing the 'rates' at which spikes would be emitted are continuously exchanged instead. However, neuroscientists believe that large amounts of information is transmitted in the precise times at which spikes are produced. Artificial 'spiking neural networks' can harness these properties, making them useful in applications which are challenging for current models such as real-world robotics and processing data with a temporal component, such as video. However, spiking neural networks can only be used effectively if suitable computer hardware and software is available. While there is existing software for simulating spiking neural networks, it has mostly been designed for studying real brains, rather than building AI systems. In this project, I am going to build a new software package which bridges this gap. It will use abstractions and processes familiar to machine learning researchers, but with techniques developed for brain simulation, allowing exciting new SNN models to be used by AI researchers. We will also explore how spiking models can be used with a special new type of sensors which directly outputs spikes rather than a stream of images. In the first phase of the project, I will focus on using Graphics Processing Units to accelerate spiking neuron networks. These devices were originally developed to speed up 3D games but have evolved into general purpose devices, widely used to accelerate scientific and AI applications. However, while these devices have become incredibly powerful and are well-suited to processing lots of data simultaneously, they are less suited to 'live' applications such as when video must be processed as fast as possible. In these situations, Field Programmable Gate Arrays - devices where the hardware itself can be re-programmed - can be significantly faster and are already being used behind the scenes in data centres. In this project, by incorporating support for FPGAs into our new software, we will make these devices more accessible to AI researchers and unlock new possibilities of using biologically-inspired spiking neural networks to learn in real-time.As well as working on these new research strands, I will also dedicate time during my fellowship to advocate for research software engineering as a valuable component of academic institutions, both via knowledge exchange and research funding. In the shorter term, I will work to develop a community of researchers involved in writing software at Sussex by organising an informal monthly 'surgery' as well as delivering specialised training on programming Graphics Processing Units and more fundamental computational and programming training for new PhD students. Finally, I will develop internship and career development opportunities for undergraduate students, to gain experience in research software engineering.