论文部分内容阅读
分析了前向神经网络误差函数的拟凸性质及其分片表示结构，提出了分块变尺度算法。该算法具有分块误差向后传播的特点。数值结果表明，算法具有良好的收敛性质
The quasi-convexity property of the error function of feedforward neural network and its fragment representation structure are analyzed. A block-based scaling algorithm is proposed. The algorithm has the characteristics of block error propagation backwards. The numerical results show that the algorithm has a good convergence property