基于Σ-Δ调制的比特流Sigmoid函数的实现及其在3-D空间判别网络中的应用  

Implementation of Sigmoid Activation Function for Sigma-Delta Modulated Bit-Streams and Its Application on Network for 3-D Space Classification

作　　者：郭晓丹[1,2] 孟桥[2] 梁勇[2] 

机构地区：[1]南京信息工程大学电子与信息学院,江苏南京210044 [2]东南大学射频与光电集成电路研究所,江苏南京210096

出　　处：《电子学报》2015年第5期862-869,共8页Acta Electronica Sinica

基　　金：国家自然科学基金(No.60576028)

摘　　要：本文提出一种基于Σ-Δ调制的比特流Tangent-Sigmoid(Tan-Sig)函数的数字电路实现方法.主要是通过修改反馈系数构建Σ-Δ限幅放大调制器,并以此组合实现对Tan-Sig函数的逼近.根据上述方法,在现场可编程门阵列上设计了具有非线性激活函数的比特流神经元以及复杂的比特流前馈型人工神经网络.最后利用该网络实现了在笛卡尔坐标系下的3-D空间判别.A digital hardware implementation of Tangent Sigrnoid （Tan-Sig） activation function based on sigma-delta modu- lated bit-streams is proposed. Through the change of feedback coefficient in the framework of traditional sigma-delta modulator, a new limiting amplifier modulator （LAM） is fabricated, and the function approximation to Tan-Sig was achieved by the combination of several LAMs with different coefficients. Meanwhile, the bit-stream neurons with Tan-Sig activation function and the whole feed- forward artificial neural networks are implemented on field programmable gate array （FPGA）. And the 3-D space classification problem in Cartesian coordinate system was solved by the neural networks presented.

关 键 词：比特流 人工神经网络 激活函数:Σ-Δ调制 

分 类 号：TN911[电子电信—通信与信息系统]