基于Σ-Δ调制的比特流Hopfield人工神经网络及在片训练系统  

Hopfield Neural Network Based on Σ-Δ Modulation and On-chip Training System

作　　者：郭晓丹[1,2] 孟桥[1,2] 

机构地区：[1]东南大学射频与光电集成电路研究所,江苏南京210096 [2]东南大学射频集成电路与系统教育部工程研究中心,江苏南京210096

出　　处：《信息与控制》2014年第1期1-8,共8页Information and Control

基　　金：国家自然科学基金资助项目(61076118/F040805)

摘　　要：提出了一种基于Σ-Δ调制的单比特Hopfield人工神经网络(HNN)的设计方法.网络中各神经元的输入输出均为基于Σ-Δ调制的单比特信号,这解决了多比特并行分布网络扇入扇出的难题.同时在设计12神经元全并行Hopfield网络过程中,通过数字后端设计,优化了关键模块,减小了系统的面积及功耗.最后采用同步扰动学习算法(SLA)实现了在片训练系统,通过此训练系统Hopfield网络实现了文字的联想记忆.A design method for the bit-stream Hopfield neural network （HNN） based on ∑-△ modulation is presen- ted. The signals from the input and output of each neuron are represented by ∑-△ modulated single-bit streams. This single-bit representation alleviates the fan-in and fan-out issues typical of distributed systems. A parallel distributed network with 12 neurons is designed, and the whole HNN is implemented on a field pro- grammable gate array and hackend design. Several key modules of the systems are optimized and the area, power are reduced by using digital backend design. Moreover, the simultaneous learning algorithm （SLA） is used to train the HNN. The SLA is an on-chip learning algorithm and is implemented on board. The bit- stream HNN achieves rather precise character recognition and memorization after the training.

关 键 词：比特流 人工神经网络 HOPFIELD ∑-△ FPGA Astro 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]