FPGA-based Fault-injection and Data Acquisition of Self-repairing Spiking Neural Network Hardware
FPGA-based Fault-injection and Data Acquisition of Self-repairing Spiking Neural Network Hardware
复制标题
DOI：
10.1109/iscas.2018.8351512
复制
发表时间：
2018-05
期刊：
影响因子：
--
通讯作者：
Shvan Karim;J. Harkin;L. McDaid;B. Gardiner;Junxiu Liu;D. Halliday;A. Tyrrell;J. Timmis;Alan G. Millard;Anju P. Johnson
中科院分区：
文献类型：
--
作者：
Shvan Karim;J. Harkin;L. McDaid;B. Gardiner;Junxiu Liu;D. Halliday;A. Tyrrell;J. Timmis;Alan G. Millard;Anju P. Johnson
尖峰星形胶质细胞神经元网络 (SANN) 模拟人脑的适应性/修复功能。它们将星形胶质细胞与尖峰神经元整合在一起，以促进突触水平上的分布式和细粒度的自我修复能力。由于添加了星形胶质细胞，SANN 变得更加复杂，并且需要更长的模拟时间，因为它们的动态时间范围比传统神经网络长得多。因此，专用 FPGA 加速器可以缩短仿真时间。为了支持 SANN 的加速，需要具有向突触进行故障注入的能力以及监控神经元和星形胶质细胞数据的重要水平，以便在片外传输到基于 PC 的分析。本文提出了一种基于 FPGA 的监控平台 (FMP)，用于注入故障以及捕获和分析从 SANN FPGA 加速器 Astrobyte 获取的数据。 FMP 使用定制逻辑和基于 NIOS II 的系统来控制 FPGA 上的故障注入和数据监控。结果显示，使用 FMP 对故障注入场景进行了准确的加速模拟，与同等的 Matlab 实现相比，速度提高了 65 倍。
Spiking Astrocyte-neuron Networks (SANNs) model the adaptive/repair feature of the human brain. They integrate astrocyte cells with spiking neurons to facilitate a distributed and fine-grained self-repair capability at the synapse level. SANNs are more complex with the addition of astrocyte cells and require longer simulation times, as they are dynamic over much longer time-scales than traditional neural networks. Therefore, dedicated FPGA accelerators offer reductions in simulation times. To support the acceleration of SANNs, the capability of fault injection to synapses and monitoring significant levels of neuron and astrocyte data for off-chip transmission to PC-based analysis, are required. This paper presents an FPGA-based monitoring platform (FMP) for injecting faults and capturing and analyzing data acquired from the SANN FPGA accelerator, Astrobyte. The FMP uses custom logic and a NIOS II based system to control fault injection and data monitoring on the FPGA. Results show accurate accelerated simulations of fault injection scenarios using FMP with speedups up to 65 times greater compared with equivalent Matlab implementations.