基于脉冲序列编码的递归脉冲神经网络深度学习方法研究

Motivated by recent advances in neurosciences, neural information is encoded in the brain through the precisely timed spike trains, not only through the neural firing rate. By mimicking the hierarchical structure of human brain, deep recurrent spiking neural networks can extract features from lower level to higher level gradually, and improve the performance for the processing of spatio-temporal information. Due to the complex hierarchical structure and implicit nonlinear mechanism, the formulation of efficient deep learning methods for recurrent spiking neural networks based on spike train encoding is difficult and remains an important problem in the research area. The main contents include as follows: (1) Using the probability distribution of the spike train patterns, an unsupervised learning algorithm based on gradient descent of the Kullback-Leibler divergence for deep recurrent spiking neural networks will be developed, which can facilitate subsequent learning through the pre-training in neural layers. (2) Using information encoded in precisely timed spike trains and error backpropagation mechanism based on feedback alignment, a supervised learning algorithm with inner product of spike trains for deep recurrent spiking neural networks will be investigated, which can improve the capability for solving complex problems. (3) The new real-time computational models will be designed for pattern classification and multimodal fusion of spatio-temporal sequence brain data using the proposed deep learning algorithms. The aim of the project is to study deep learning algorithms for hierarchical recurrent spiking neural networks. Therefore, this work will be provided effective theoretical foundation and technical support for research on the spatio-temporal pattern recognition.

神经科学研究表明，神经信息被编码为精确定时的脉冲序列，不只是简单的脉冲发放频率。深度递归脉冲神经网络模拟人脑的多层结构，对数据从低层到高层渐进地进行特征提取，进而提高对复杂时空信息的处理能力。然而，由于递归脉冲神经网络复杂的层次结构和内在的非线性机制，构建高效的基于脉冲序列编码的深度学习方法非常困难。本项目的研究主要包括三个方面：(1)应用脉冲序列模式的概率分布，设计基于Kullback-Leibler散度梯度下降的无监督学习算法，实现自下而上逐层的预训练过程。(2)深入分析脉冲序列的信息编码特点和基于反馈比对的误差反向传播机制，设计基于脉冲序列内积表示的监督学习算法，提高对复杂问题的求解能力。(3)发展脑时空序列数据模式分类与多模态融合的深度学习实时计算模型，验证所提算法的有效性。本项目旨在创建递归脉冲神经网络的深度学习方法，为各种时空模式识别应用奠定坚实的理论基础和有力的技术支持。

神经科学研究表明，神经信息被编码为精确定时的脉冲序列，不只是简单的脉冲发放频率。深度递归脉冲神经网络模拟人脑的多层结构，对数据从低层到高层渐进地进行特征提取，进而提高对复杂时空信息的处理能力。本项目针对递归脉冲神经网络复杂的层次结构和内在的非线性机制，面向脑时空序列数据实时分类和融合问题，研究了递归脉冲神经网络的深度学习方法。（1）基于神经信息的脉冲序列编码方式，构建了具有分层结构的递归脉冲神经网络的软件模拟平台,深度递归脉冲神经网络由多个递归脉冲神经机模块堆叠构成。（2）应用脉冲序列模式的概率分布，提出了基于脉冲驱动的深度递归脉冲神经网络无监督学习算法，实现自下而上逐层的预训练过程，相应的脉冲序列学习实验表明所提算法具有较低的重建误差。（3）基于反馈比对的误差反向传播机制，提出了深度递归脉冲神经网络的监督学习算法，网络中相邻层之间的前馈连接与层内侧向的递归连接权值的学习规则表示为脉冲序列内积的形式，实现了脉冲序列复杂时空模式的学习能力。将所提算法应用于脉冲序列的学习任务和非线性模式分类问题，取得了较高的学习精度和良好的适应能力。此外，基于脉冲序列内积表示的学习规则使所提监督学习算法具有更广的适用性，可用不同的脉冲神经元模型构建深度递归脉冲神经网络的学习模型。（4）将脑时空序列数据（如脑电图、功能性核磁共振成像等）编码为脉冲序列，构建了脑时空序列数据模式分类与多模态融合的深度学习实时计算模型，首先应用无监督学习算法进行逐层预训练，然后应用监督学习算法反向微调整个网络，提高了深度递归脉冲神经网络对复杂问题的实时处理能力。实验中，通过癫痫等神经疾病的脑时空序列数据集验证了所提模型的有效性。