摘要
针对自组织映射(SOM)神经网络算法实现复杂的问题,提出SOM算法的简化方案及并行硬件电路架构.经典SOM算法中,权值更新函数须使用浮点数乘法、开方以及指数等运算,硬件并行实现十分困难.传统的SOM简化方法的聚类准确率不高,面对手写数字识别这类复杂应用,传统方法的识别率十分有限.提出的SOM简化算法可以在保证系统聚类准确率的同时,除去权值更新函数中的复杂运算,易于硬件的全并行实现.基于提出的SOM简化算法及并行电路架构,实现的手写数字识别系统的工作频率为50 MHz,单次输入的学习时间仅需200ns,实时处理性能可达400MCUPS.识别系统针对MNIST样本库的识别准确率超过81%,与经典SOM算法的准确率持平,明显优于其他SOM简化方法.
A simplified self-organizing map （SOM） algorithm and its parallel hardware architecture were proposed in order to handle the complex problem of hardware implementation of SOM. The weight update function of the conventional SOM contains multiplication, square and exponential operations, which makes parallel realization difficult. Traditional simplified SOM methods have low accuracy on classification, especially when encountering complex applications such as handwritten digit recognition. The proposed SOM algorithm has reduced all complex computations without sacrificing accuracy, and full parallel hardware implementation is possible. The proposed hardware system can proceed at 50 MHz and achieve a performance of 400 MCUPS, which means that learning a single input pattern will take only 200 ns of time. When applying the handwritten digit recognition on the MNIST database, the proposed system can recognize over 81% of the patterns correctly, which is almost the same accuracy as the conventional SOM, but is much better than other simplified SOM methods.
出处
《浙江大学学报（工学版）》
EI
CAS
CSCD
北大核心
2014年第4期742-747,共6页
Journal of Zhejiang University：Engineering Science
基金
国家自然科学基金资助项目(61204030)
浙江省自然科学基金资助项目(LY13F020030
LQ12F04002)
浙江省重中之重学科资助项目
