可继续学习的构造型神经网络构造算法

The Design of Incremental Learning for Constructive NN

构造型神经网络模型通过将样本映射到单位超球面上并用覆盖方法进行识别,具有计算速度快、识别率高、几何意义明显等优点。但是常用的基于交叉覆盖的方法在首次构造完成后,难以再继续进行修改和加强,从而阻碍了网络的再学习能力。文章提出了该构造型神经网络的一种双交叉覆盖方法,一方面吸收了原交叉覆盖的优点,一方面提供了良好的再学习能力。通过实验验证,该方法可以较好地运用到构造型神经网络的增量学习中。

The constructive neural network,which established a mapping from original sample to a super sphere,could identify the different kinds of samples quickly and accurately.But it was hard to be modified and enhanced after the first conformation.In this paper,a new incremental learning algorithm is proposed,it is fit for incremental learning after the first conformation and also has the advantages of the constructive neural network.The experiment demonstrate that the new algorithm(BiCA)has the strong learning ability and can obtain the new knowledge from the sample quickly and steady.