任务自适应神经网络结构研究

A Study of Task Based Strategies for Adaptively Constructive Neural Networks

在基于任务自适应的统一框架下讨论了神经网络的生成方案，重点研究了两类典型而适用的神经网络构筑算法，即基于隐节点自适应增长的神经网络结构算法和基于子网自适应增长的神经网络结构算法。还结合所提出的层次式多网络模型，对基于任务的神经网络的自适应结构方案、发展前景与存在的问题作了详细的比较研究。

We investigated the strategies for optimizing neural networks under the unified frame based on task, focused for constructive neural networks on two typical and practical schemes, which are adaptively constructive neural networks by growing hiddens or layers of hidden nodes and by growing subnet. With the Layer Multinet Model proposed by our research group, we investigated task based algorithms for constructive neural networks, their perspective, strength and weakness.