Research Of The Algorithms Of The Spiking Neural Networks

Deep learning,a major driving force behind artificial intelligence(AI),has achieved remarkable progress in the fields of image,vision and speech processing,in recent years.However,there is still an unbridgeable gap between deep learning and human-level intelligence or general purpose AI.To realize this lofty goal,neuromorphic computing provides an alternative attempt.In the core of neuromorphic computing lays Spiking Neural Networks(SNNs)that is modelled after the spiking mechanism similar to a biological neuron.SNNs are able to process and extract features from the temporal dynamics encoded in the spike signals,unlike the traditional frequency-based neural networks,thus making SNNs more biologically plausible and also easier to deploy on hardware.However,SNNs cannot be learnt using the existing algorithms applicable to learning a neural network due to the complexity in encoding and the undifferentiability of the spiking variables.Right now,there lacks an efficient learning algorithm for SNNs,which is the reason for its limited use.This thesis centres on the topic of SNNs learning algorithms;analyzes the strengths and weaknesses of the existing algorithms and explores robust and efficient learning algorithms for SNNs.The thesis makes the following contributions:1.Firstly,this paper proposes an effective and robust membrane potential-driven supervised learning(Mem Po-Learn)method.One of the fundamental computations of spiking neurons is to transform streams of input spike trains into precisely timed firing activity.While the traditional spike-driven learning methods use an error function based on the difference between the actual and desired output spike trains,the proposed Mem Po-Learn method employs an error function based on the difference between the output neuron membrane potential and its firing threshold.The experimental results demonstrate that the proposed Mem Po-Learn enables the trained neurons to generate desired spike trains with higher precision and higher efficiency than the current state-of-the-art spiking neuron learning methods.2.Secondly,this paper proposes a skip scan training strategy(SSTS),for learning precise time-step spikes.Since small time steps can be approximated to continuous time and it is extremely important for real-time applications of SNNs.Therefore,we start with analyzing the effect of using a small time step on the learning complexity.Using a smaller time step in learning is more difficult and time-consuming.Hence,SSTS was proposed to improve the efficiency of small time-step based learning.Experimental results demonstrate that SSTS significantly improves the efficiency of the Mem Po-Learn method.3.Thirdly,this paper proposes a noise-threshold to improve the noise tolerance of spiking neurons with a dynamic firing threshold during training.Noise is prevalent in spiking neural networks and it often perturbs the neural response of the system.The proposed noise threshold draws strength from a thorough analysis of the existing strategies combined with the inspiration from biological neurons to make the spiking neurons robust to noise.The noise-threshold can be applied by the existing supervised learning methods to improve their noise tolerance.Experimental results show that,with noise-threshold,the anti-noise capability of the existing supervised learning methods improves significantly.4.Lastly,this paper proposes a novel membrane-potential driven aggregate-label learning algorithm,namely MPD-AL.One of the long-standing questions in biology and machine learning is how neural networks could learn important features from the input activities with delayed feedback.This is commonly known as the temporal credit-assignment problem.Aggregate-label learning is proposed to solve this problem.However,the existing threshold driven aggregate-label learning algorithms is computationally expensive.This low learning efficiency limits their usability in practical applications.In order to address these limitations,we propose a novel membrane-potential driven aggregate-label learning algorithm.The experimental results demonstrate that the learning performance of the proposed MPD-AL is better than the existing learning algorithms.