一种定点数权值神经网络的优化方法及其应用

为了使得神经网络的应用符合嵌入式系统快速计算、存储量精简的要求，提出了一种定点数权值神经网络的优化方法。采用精度可调的比例数格式定点数表示神经网络的权值和阈值，用遗传算法对神经网络进行训练，并用最小二乘法对网络的非线性连续节点激励函数进行了线性离散化。将这种优化的神经网络应用于触摸屏校准。实验表明，采用该方法进行触摸屏校准比传统的校准方法具有更高的准确率。

Optimization of neural network with fixed-point number weights and its application

In order to make the neural network application suffice the demands of double-quick computing and tidy memory capacitance in embedded systems, an optimization method of neural network with fixed-point number was proposed. The neural network weights were represented with the precision-adjustable fixed-point number and the neural network was trained by using the genetic algorithm. And the continuous nonlinear activation function of the neuron was transformed into discrete and linear function by the least-squares algortithm. Then, the optimal neural network was applied to a touch-screen-LCD adjusting model for verifying its feasibility. Experiments show that this touch-screen-LCD calibration method has higher accuracy compared with the traditional one.