论文部分内容阅读
为了避免奇异解,提高网络性能,给出一种回声状态网络的权值初始化方法(WIESN).利用柯西不等式和线性代数确定优化的初始权值的范围与输入维数、储备池维数、输入变量和储备池状态相关,从而确保神经元的输出位于sigmoid函数的激活区域.实验结果表明,权值初始化方法的精度和训练时间要优于随机初始化方法,且相比于训练时间,权值初始化的时间是可以忽略不计的.
In order to avoid the singular solution and improve the network performance, a WIESN method is proposed.With Cauchy inequality and linear algebra, the range of initial weights and the number of input dimensions, reserve pool dimension, The input variable is related to the state of the reserve pool to ensure that the output of the neuron is located in the activation region of the sigmoid function.The experimental results show that the precision and training time of the weight initialization method is better than that of the random initialization method and compared with the training time, The initialization time is negligible.