前馈神经网络隐层评测问题的研究

Study on Hidden Layer Evaluation for Feed-Forward Neural Networks

分析了隐层输出向量组生成的表示空间与期望输出向量组生成的目标空间.通过计算隐单元的误差补偿值,对以隐层生长方式构建网络时,每个隐单元的误差补偿性能以及隐单元性能最优的充分必要条件进行了研究.结果表明:表示空间与目标空闽维数、隐单元数目以及每个隐单元的误差补偿效率决定了前馈神经网络隐层的评测因素.最后定义了隐层品质因子、隐层有效系数、隐单元剩余度和隐层评价因子,并通过对典型前馈网络的考察,验证了该评测方法的合理性和有效性.

The representation space and the target space, subspaces generated respectively by output vectors of hidden layer and expected output vectors, were analyzed. For hidden-layer-growing construction method, the error compensation performance of hidden units and the sufficient and necessary condition for the optimal performance of hidden units were investigated by calculating their error compensation values. The investigation shows that the evaluation factors for the hidden layer of a feed-forward neural network are composed of the dimensions of representation space and target space, the number of hidden units, and the error compensation efficiency of each hidden unit. Finally, the quality factor of hidden layer, the efficient coefficient of hidden layer, the redundancy of hidden units and the evaluation factor of hidden layer were defined, and the rationality and validity of the evaluation method were verified by reviewing some typical feedforward neural networks.