一种基于FPGA的硬件神经网络设计方法

Neural Networks Hardware Implementation Method Based on FPGA

在FPGA硬件神经网络设计中激活函数的实现和数据表示方式是两个难点。本文提出了用非线性函数和21位定点法相结合来实现激活函数的逼近算法,采用源码定点表示法实现数据的硬件表示,明显减少了FPGA的资源占用,降低了激活函数逼近算法的复杂性和实现难度,最后,给出实际FPGA硬件神经网络设计实例并进行了仿真验证。

The most difficult problems encountered when implementing artificial neural networks based FPGA are the approximation of the activation function and the data representation.It presents the non-linear approximation of the activation function and the presentation of 21 bit fixed-point decimal to resolve the problems above.The methods achieve require obviously less hardware resources and decrease the difficulty of approximation algorithm the activation function and the implementation of it.At last,give an example of the implementation of a real Artificial Neural Networks using FPGA and its Simulation Verification.