FPGA-based Reservoir Computing with Optimized Reservoir Node Architecture
FPGA-based Reservoir Computing with Optimized Reservoir Node Architecture
复制标题
基于FPGA的油藏计算，优化油藏节点架构
DOI：
10.1109/isqed54688.2022.9806247
复制
发表时间：
2022
期刊：
影响因子：
--
通讯作者：
Yi, Yang
中科院分区：
文献类型：
--
作者：
Lin, Chunxiao;Liang, Yibin;Yi, Yang
更新水库节点的状态是水库计算（RC）的基本操作之一，它极大地影响系统的性能。在回声状态网络（ESN）（RC 的主要类型之一）中，状态更新过程可分为两个阶段：权重矩阵与输入状态向量相乘并对乘积之和应用非线性激活函数。权重矩阵通常较大且稀疏，为优化矩阵乘法提供了机会；激活函数的选择也可能影响硬件资源的利用率。本文介绍了一种基于 FPGA 的 RC 系统的优化储层节点架构。具体来说，我们采用位串行矩阵乘法器和权重矩阵的直接空间实现来充分利用稀疏性。规范的有符号数字表示也用于进一步优化乘法器逻辑。此外，设计并优化了双曲正切激活函数，以保持神经网络的非线性而不影响其准确性。与现有的硬件ESN设计相比，我们的水库节点架构显着降低了资源利用率，同时保持了可比的性能。
Updating the state of reservoir nodes is one of the essential operations of reservoir computing (RC), which highly affects the system’s performance. In an echo state network (ESN), one of the primary types of RC, the process of state renewal can be divided into two stages: multiplication of the weight matrix with the input-state vector and applying a nonlinear activation function on the sum of products. The weight matrix is typically large and sparse, providing opportunities for optimizing the matrix multiplication; the choices of activation functions may also affect hardware resource utilization. This paper introduces an optimized reservoir node architecture for FPGA-based RC systems. Specifically, we adopt the bit-serial matrix multiplier and direct spatial implementation of the weight matrix to fully exploit the sparseness property. The canonical signed digit representation is also employed to further optimize the multiplier logic. Furthermore, a hyperbolic tangent activation function is designed and optimized to maintain the nonlinearity of the neural network without affecting its accuracy. Compared with existing hardware ESN designs, our reservoir node architecture significantly reduces resource utilization while maintaining comparable performance.
神经形态计算处理器中基于 FPGA 的尖峰时间相关编码器和储存器设计
影响因子：
2.6
作者：
Yang Yi;Y. Liao;Bin Wang;Xin Fu;Fangyang Shen;Hongyan Hou;Lingjia Liu
通讯作者：
Lingjia Liu
影响因子：
2.2
作者：
Bai, Kangjun;Yi, Yang
通讯作者：
Yi, Yang
在时间序列预测的分解方法中使用储层计算
DOI：
10.1016/j.micpro.2016.03.009
发表时间：
2008
期刊：
Microprocess. Microsystems
影响因子：
--
作者：
F. Wyffels;B. Schrauwen;D. Stroobandt
通讯作者：
D. Stroobandt