Spiking神经网络学习算法研究

Spiking neuron model is widely studied in biological sciences and neuroscience. Whereas,few success have been reported to solve real world problems. One important reason for this is that the current learning algorithms for spiking neural networks are too limited. Today much evidence has been revealed in neuroscience that learning in biological neural networks is correlation-based. However cognitive behavior is often considered to be target-driven, which indicates a supervised approach to learning rather than pure correlation-based learning. Whereas some learning algorithms have been developed for spiking neural networks, most of those are still correlation-based and difficult to solve real-world problems. This project attempts to research the learning algorithms for spiking neural networks based on some recent achievements of neuroscience and cognitive science by using machine learning ideas. More specifically, we analyze the temporal-coding mechanism of spiking neural networks, study biologically plausible spiking learning algorithms with experimentally supported learning rules (Correlation-Based) and recognized cognitive behaviors(Target-Driven), and deliver a series of efficient learning algorithms for spiking neural networks that are able to learn tasks in a goal-oriented, target-driven way, and not tied to a specific task but able to learn arbitrary input-output relationships. The research can contribute to understanding the learning process of cognitive science, computational neuroscience and machine learning in a high degree.

神经科学揭示生物神经元学习过程是Correlation-Based，而生物的认知行为通常是Target-Driven。对Spiking神经网络学习算法的研究，目前多数是借鉴生物correlation-based学习特性，缺少对认知行为的模拟，存在不少缺陷，很难用于解决实际工程问题。本课题在分析神经科学、认知科学最新成果基础上，利用机器学习理论方法，研究Spiking神经网络的学习算法。具体而言，是分析Spiking神经网络时间编码机制，综合考虑神经系统Correlation-Based学习特性和认知行为Target-Driven学习常规，利用符合认知行为的监督学习和强化学习策略，研究既具有高度仿生物性能、又符合认知行为规则,有广泛适用范围的以目标为向导、以目标为驱动、不依赖特定任务和特定输入输出的高效Spiking神经网络学习算法。从而使认知科学、神经科学和机器学习达到高度统一。

本课题在神经科学、认知科学最新研究成果基础上,利用机器学习理论方法,研究Spiking神经网络的学习算法。研究探索既有高度仿生物性能,又符合生物认知行为规则,有广泛适用范围的以目标为向导、以目标为驱动、不依赖特定任务和特定输入输出的高效Spiking神经网络学习算法。本课题研究Spiking神经网络学习算法，完成情况概述如下：.（1）通过分析现有Spike-driven算法的局限性，在该领域率先提出Membrane potential-driven的学习策略。Membrane potential-driven学习策略不仅符合生物认知行为规则，而且显著的提高了Spiking神经网络学习的准确性和高效性。.（2）为了提高Spiking神经网络学习算法的鲁棒性，首次引入了动态阈值的权重调整策略。提高了Spiking神经网络学习算法的鲁棒性。.（3）利用机器学习的算法，提出一种可以训练深度spiking神经网络的学习算法。.（4）提出一种基于spiking神经网络的图像特征提取方法，并应用于模式识别和人体行为识别。.（5）利用我们建立的具有波传递特性的Spiking神经网络，提出一种适合在复杂环境中应用的最短路径计算方法。.在项目的执行过程中，我们严格按照计划，完全了预定的任务，取得的研究成果总结如下：.（1）在国际一流刊物 IEEE Transaction 上发表和录用学术论文6篇。.（2）发表和录用科研论文20篇，其中SCI检索12篇。.（3）项目负责人申请中国技术发明专利7项，其中已授权2项。.（4）通过本课题的研究，我们显著的提高了Spiking神经网络学习算法的准确性，高效性和鲁棒性。其中，学习算法的效率提高了近两个数量级。.（5） 获得教育部自然科学奖一等奖。.Spiking神经网络学习方法,是一种与认知科学、神经计算和机器学习相互交叉的前言性研究课题,是对生物认知系统和生物神经系统庞大功能的模拟和探索,能进一步加强人们对大脑如何进行记忆、识别和认知的理解,具有极高的科学意义。