Abstract: Presented a novel self-architecure modular neural network architecture, called modular-tree for supervised learning. In the architecture, any kind of feedforward neural networks can be employed as componenets and a modular neural network with the tree structure is generated automatically with a growing algorithm by partitioning input space recursively to avoid the problem of pre-determined structure. Due to the principle of divide-and-conquer used in the proposed architecture, the modular-tree can yield both a good performance and significantly fast training. The proposed architecture has been applied to several supervised learning tasks including both benchmark and real-world problems and achieved satisfactory results.

Key words: modular neural networks, self-architecture, supervised learning

摘要： 提出了一种新颖的具有自构筑能力的神经网络结构，称之为Modular-tree和两个相应的自构筑算法。在此结构中，任何现存的前馈神经网络均可以作为子网。对于一个给定的学习任务，利用提出的生成算法通过对输入空间递归地划分，自动生成一树状的模块神经网络，从而避免了网络结构预置问题。由于使用了“分治”原理，Modular-tree具有良好的性能及快速训练的能力。此结构已用于多个监督学习问题(包括：标准测试及现实世界问题)并取得令人满意的实验结果。

关键词: 模块神经网络, 自构筑, 监督学习