激励函数可学习神经网络

Active Functions Learning Neural Network

提出了一种激励函数可学习神经网络,其神经元函数不固定,通常是任何线性无关的基函数的线性组合,通过调整神经元中基函数的系数即可达到网络学习的目的。为了结构优化方便,将神经元输出的多维空间映射为一维空间后输入给下层神经元。根据网络的特点,提出了两种无需迭代的网络参数快速学习算法实现网络训练。通过3个实例进行仿真实验,结果表明所设计神经网络的逼近能力强,参数学习速度极快。

In this paper,a novel neural network is proposed,whose active functions can be learned. Its active functions are not given and cannot be changed,but can be learned by the problems and could be the linear combination of any linear independent basis functions. The networks could be learned by tuning the coefficients of the basis functions. For the convenience of structure optimization,the input vectors of neurons are always mapped to be scalar variables. In this paper,two quick learning algorithms are proposed for this network which does not need iterative procedures. The results show that this network has good function approximation capacity and fast learning speed.