【作者】 孙伟；

【导师】 于强； 刘登科；

【作者基本信息】 天津大学 ， 电子信息， 2023， 硕士

【摘要】 近年来,以深度学习为代表的人工神经网络获得飞速发展,并在众多感知任务中取得了超过人类识别准确率的性能表现。然而基于深度学习的方法往往面临对大量的训练数据与巨大的计算资源需求的发展瓶颈。考虑到人脑的低功耗、强大的认知与学习能力,受脑运算机制启发的脉冲神经网络为人工神经网络的未来发展提供了新的思路,近年来受到越来越多的关注。脉冲神经网络利用脉冲来表示和传递信息,是类脑和神经形态计算领域的主要研究内容。最近有一系列多脉冲学习算法被提出,用于处理和学习脉冲序列中携带的信息,并在各种模式识别任务中取得了良好的效果,但是目前这些基于脉冲的学习算法都没有考虑输入信息的时间波动,而输入时间波动是语音等感觉信号变化的常见来源。因此,在产生时间压缩或扩张的情况下,基于脉冲的鲁棒性学习具有重要的意义。本文围绕多脉冲学习算法这一主题,探索如何有效处理产生时间波动干扰的信息,提出两种高效多脉冲学习算法,本文的主要贡献如下:(1)本文结合基于电导的LIF神经元膜电位的时间扭曲不变机制,提出了面向时间扭曲不变的多脉冲学习算法(Time-warp-invariant Multi-spike Learning,TWI-ML)。为了解决该神经元膜电位解析解计算复杂度较高的问题,本文引入了基于时间步长的仿真方式。本文探究了多个参数的初始设置对时间扭曲不变性和算法性能的影响,评估了算法对不同编码机制的适应性及信息时间扭曲时的性能表现,验证了算法检测背景活动中特征的能力,以及当特征被时间扭曲时的适应性。在语音识别任务上的实验结果表明,TWI-ML算法与其他多脉冲学习算法相比具有更好的性能。此外,TWI-ML算法在处理有速率波动的语音识别方面具有更好的鲁棒性,这验证了TWI-ML算法的有效性和可行性。(2)为了进一步提高学习能力和提升计算效率,本文对基于电导的LIF神经元模型进行了改进,使用权重代替了突触峰值电导,消除了权重的非负约束,减少了反转电位的转换。基于改进的神经元模型,本文提出了新的多脉冲学习算法Improved TWI-ML。实验结果表明,与TWI-ML算法相比,Improved TWI-ML算法提高了学习能力和计算效率,增强了神经元的分类能力。在语音识别任务上,Improved TWI-ML算法比TWI-ML算法的识别准确率更高,这表明改进的算法不仅高效,而且具有更好的性能。综上,本文结合基于电导的LIF神经元模型及其膜电位动态的时间扭曲不变性,对多脉冲学习算法进行了扩展和研究,提出的多脉冲学习算法可以对时间维度受波动干扰的信息进行有效地处理,提升了学习的鲁棒性能,扩展了基于脉冲处理和学习的新范围。更多还原

【Abstract】 In recent years,artificial neural networks represented by deep learning have de-veloped rapidly,which have achieved performance exceeding the accuracy of human recognition in many perception tasks.However,methods based on deep learning often face the development bottleneck of a large amount of training data and huge computing resource requirements.Considering the low power consumption and powerful cognition and learning ability of the human brain,the spiking neural network inspired by the brain computing mechanism provides a new idea for the future development of artificial neu-ral networks,and has received more and more attention in recent years.Spiking neural networks use spikes to represent and transmit information,which is the main research content in the field of brain-like and neuromorphic computing.Recently,a series of multi-spike learning algorithms have been proposed to process and learn the informa-tion carried in the spike train,and have achieved good results in various pattern recog-nition tasks.However,most of the current spike-based learning methods are developed without considerations of input temporal fluctuations that constitute a common source of variability in sensory signals such as speech.Therefore,robust spike-based learning under fluctuations of both compression and dilation remains intriguing for exploration.Focusing on the topic of multi-spike learning algorithms,this paper explores how to effectively process information that generates temporal fluctuations,and proposes two efficient multi-spike learning algorithms.The main contributions of this paper are as follows:(1)In this paper,a time-warp-invariant multi-spike Learning(TWI-ML)algo-rithm is proposed by combining the time-warp invariance of the membrane potential of conductance-based LIF neuron.In order to solve the problem of high computational complexity of the analytical solution of the neuron membrane potential,a simulation based on time step is introduced.Then this paper explores the impact of initial settings of parameters on time-warp invariance and performance,evaluates the adaptability of the proposed algorithm to different encoding mechanisms and performance when spike information is time warped,and verifies the ability to detect features from background activity,and the adaptation when features are time warped.Experiment results on the speech recognition show that TWI-ML has better performance than other multi-spike learning algorithms.Moreover,our method has better robustness for speech recogni-tion with rate fluctuations,which verifies the effectiveness and feasibility.(2)In order to improve the learning ability and computational efficiency,this paper improves the conductance-based LIF neuron model by using weights instead of synap-tic peak conductance,eliminating the non-negative constraints of weights,and reducing the conversion of reversal potentials.Based on the improved neuron model,this paper proposes a new multi-spike learning algorithm Improved TWI-ML.Experimental re-sults show that,compared with the TWI-ML,Improved TWI-ML improves the learning ability and calculation efficiency,and enhances the classification ability of neurons.In the task of speech recognition,Improved TWI-ML has higher recognition accuracy than TWI-ML,which shows that the proposed algorithm is not only efficient,but also has better performance.In summary,this paper extends the study of multi-spike learning algorithm com-bined with the conductance-based LIF neuron model and the time-warp invariance of membrane potential.The proposed multi-spike learning algorithms can effectively pro-cess the temporal fluctuation interference of information,and improve the robust per-formance of learning,extending a new scope for spike-based processing and learning.更多还原