一种新型的神经元激活函数及其导数的可编程发生器

在神经网络的片上学习中 ,需要将神经元激活函数及其导数都映射为硬件。为满足这种需求 ,利用前向差分的原理 ,提出了一种新型的神经元激活函数及其导数的可编程发生器。该发生器采用模拟电路 ,具有结构简单、速度快、功耗小的优点。它产生的函数不仅与理想函数的拟合程度很好 ,而且可对阈值和增益因子进行编程 ,从而克服了一般模拟电路可编程性差、适用范围窄的缺点。采用标准 1 .2μm CMOS工艺的第 47级模型 ,对电路进行的 HSPICE模拟结果表明该发生器性能优越

A Novel Programmable Generator of Neuron Activation Function and Its Derivative

In the neural network on chip learning,both the neuron activation function and its derivative must be mapped to hardware. A novel programmable generator is presented to meet this requirement. Using the forward difference method, the neuron activation function and its derivative are generated at the same time. Designed by the analog circuit, the generator has the merits of simplicity, speediness and low power consumption. The generated functions are not only fit well for the ideal functions, but also programmable with the threshold and the gain factor. Thus, the generator overcomes the disadvatages of poor programmability and small flexibility of analog circuits. The circuit is simulated with HSPICE, using level 47 transistor models for a standard 1.2 μm CMOS process. The results show the superior performance of the generator.