可编程双极性sigmoid函数及其导数发生器

Programmable generator of ambipolar Sigmoid function and its derivative

在硬件实现的神经网络中,为了解决神经元激活电路结构复杂、参数不可调、与突触电路不匹配的问题,基于跨导线性原理,提出了一种双极性sigmoid激活函数及其导数发生器。该函数发生器仅由一个跨导线性环和两个差分跨导电路构成,通过改变外部偏置电流和电压可调节函数的幅值、阈值和增益因子,具有结构简单、易于编程的优点。实现的电路可选择电流或者电压输出,适用于各种电流模式或者电压模式神经网络中。采用TSMC 0.35um CMOS集成工艺对电路进行PSPICE仿真测试,结果表明,该电路产生的函数及其导数与理想函数拟合度好,相对误分别不超过2%和4%。

In the hardware realized neural network, in order to solve the problems of activation function circuit, such as complicated structure, unadjustable parameter and mismatch with the synapse, an ambipolor Sigmoid activation function and its derivative generator based on the tanslinear principle was presented. The proposed generator consists only of a translinear loop circuit and two difference transconductance circuits. The amplitude, the threshold and the gain factor of the generated function can be adjusted through external bias current or voltage, so it has the merits of simplicity and programmability. The current or the voltage can be selected as the output signal, so the generator is applicable to current-mode or voltage-mode neural networks. The proposed circuit is simulated and analyzed with PSPICE software, using TSMC 0.35um CMOS technology parameters. The results show that the proposed circuit fits well with the ideal function, the relative error of the generated function and its derivative is less than 2% and 4% respectively.