一种学习速率自适应的可编程片上学习BP神经网络电路系统的设计

设计了一种学习速率自适应的可编程片上学习BP神经网络电路系统.整个系统由前向网络、误差反传网络两部分组成.提出了一种新型的可编程S型函数及其导数的发生器电路.它不仅产生S型函数,完成非线性I－V转换;还利用前向差分法,产生S型函数的导数.这两种函数不仅与理想函数的拟合程度很好,而且易实现对阈值和增益因子的编程.为提高BP神经网络片上学习的收敛速度,还提出了学习速率自适应电路.本文采用标准1.2μm CMOS工艺的模型参数,对整个系统进行了sin(x)函数拟合等模拟实验,验证了该片上学习BP神经网络的优越性能.

A Circuit System Design of Programmable BP On-Chip Learning Neural Network with Learning Rate Adaptation

A circuit system of programmable BP on-chip learning neural network with learning rate adaptation is designed.The whole system comprises feedforward network and error back-propagation network.A novel programmable generator of sigmoidal function and its derivation is proposed.Its outputs include the sigmoidal function to realize I－V nonlinear transfer and its derivative using the forward differential method.Both functions fit well with the ideal functions.Moreover, the threshold and the gain factor can be easily programmable.Learning rate adaptation circuit is also presented to accelerate the convergent speed.Using a standard 1.2μm CMOS process, experiments such as sin(x) function fitness are done to the whole system.These experiments verify the superior performance of this on-chip learning BP neural network.