自适应加噪反向传播算法研究

提出了一种对三层前向感知器的权重修正值注入自适应噪声的方法。在误差反向传播训练过程中 ,感知器权重的噪声幅度随权梯度值下降而趋近于零。不论是仅采用权重加噪的方法还是同时进行训练集输入端及权重加噪的方法 ,都能显著提高神经网络的泛化能力 ,并以异或问题的仿真实验验证了上述结论

Study on the Adaptive Noise Injection in Back Propagation Training

In this paper a novel way of injecting adaptive noise into the weights and biases of the multilayer perceptron to improve the generalization performance is presented. During the training period, the noise level multiplied to the weights and biases is adjusted with the gradient of the cost function, and approaches zero gradually. By injecting noise into both the weights and biases of the network or the input vectors of the training sets, the generalization performance is greatly enhanced.