脑启发计算平台高效立体互连架构及其通信可靠性研究

Brain-inspired computing systems are beneficial for the computations and simulations of the brain functionalities and the developments of the intelligent hardware systems. However, current approaches did not fully address the interconnect challenges of the large-scale neuron networks and the system scalability and robust communication can be further explored. Based on the third generation of neuron network model, i.e. spiking neural network, this project aims to develop hierarchical interconnect structure and robust communication strategies for the brain-inspired computing platform. The research objectives include: to develop the optimised computational models for the key components in the neural network that will still maintain the important biological behaviours and design the corresponding hardware structure with low resource cost; to use the three dimensional Networks-on-Chip and Address-Event-Representation techniques to develop hierarchical interconnect structure by taking inspiration from the biological structure of the human brain, and to explore the mapping algorithms for the neural network topology and system task that will aid creating the communication system with high throughput and low latency; to provide novel algorithms and mechanisms to enhance the communication robustness and fault-tolerant capability of the proposed hierarchical interconnect structure. This project would address the intensive connection and robust communication problems between the large volume of neurons and synapses, and provide the computing platforms and methodology references for the data analysis and simulation of the brain functionalities, the development of the cognitive computing and learning platforms, and the design of artificial intelligent hardware systems.

脑启发计算平台对脑功能计算模拟及智能硬件研制具有重要意义,然而现有的互连策略无法有效解决大规模神经单元间的互连问题，特别在系统可扩展性以及通信鲁棒性等方面有待进一步研究。本项目将以第三代脉冲神经网络模型为基础，研究构建脑启发计算平台的立体互连模型及其通信可靠性策略。主要包括：研究设计关键神经网络生物细胞的优化计算模型及在保持其生物学特性前提下的单位节点低能耗硬件结构设计，借鉴大脑生物神经网络的层次化结构，利用三维立体片上网络技术和地址事件表达机制建立神经网络立体互连架构，并研究拓扑结构及系统任务关系映射算法来辅助构建高吞吐量、低时延的通信系统，且提出新的算法和机制以加强神经网络立体互连架构的通信可靠性和容错能力。本项目能够解决脑启发计算平台的大规模神经元、突触互连时存在的密集连接及通信可靠性问题，为脑机理的数据分析与模拟、认知学习平台的构建以及人工智能硬件系统的设计提供物理平台与方法指导。

项目研究了神经元网络关键生物细胞的计算模型优化与可扩展性，提出了多种优化模型对脉冲神经网络学习算法进行改进，从而提高其抗噪声能力和鲁棒性；对其生物神经网络计算模型进行分析及优化，降低模型计算复杂度，保障关键细胞在大规模系统实现时的可扩展性；研究了典型神经网络系统的模型优化和泛化能力，优化神经网络模型来提升在时间序列数据上的系统性能；研究设计了脉冲神经网络立体硬件互连通信架构，提出了基于片上网络技术的神经元节点立体互连通信方案，构造了类生物神经网络的层次化硬件互连结构；研究了生物神经网络和硬件系统间的任务映射算法，根据神经网络系统任务模型对节点互连的拓扑结构以及通信模型进行优化目标求解和分析，在通信时延、吞吐率以及功耗等方面取得了较好的性能；提出了脉冲神经网络硬件系统容错性与可靠性保障机制，研究设计了仿生硬件容错架构，为神经网络硬件的可靠性提供保障，构建了具有高容错性的生物神经网络电子硬件系统；采用典型测试数据集和实际应用场景对优化后的神经网络系统进行功能验证，进行了理论模型和实际应用测试分析，所提出设计的神经网络模型及硬件系统具有较好性能。在项目实施过程中，在国内外期刊发表了22篇学术论文，参加国内外学术会议交流8人次，培养硕士研究生8人。