An Incremental Neural Network for Online Supervised Learning and Topology Learning
An Incremental Neural Network for Online Supervised Learning and Topology Learning
复制标题
DOI：
10.20965/jaciii.2007.p0087
复制
发表时间：
2007-01
期刊：
影响因子：
--
通讯作者：
Y. Kamiya;S. Furao;O. Hasegawa
中科院分区：
文献类型：
--
作者：
Y. Kamiya;S. Furao;O. Hasegawa
为在线监督学习设计了一种新的自组织增量网络。在网络学习过程中，使用自适应相似度阈值来判断在线训练数据引入系统时是否需要添加新的节点。消除噪声引起的节点，减少误分类。该网络对噪声训练数据具有鲁棒性，适合于以下任务：(1)在线甚至终身监督学习；(2)增量学习，即在不破坏旧的学习信息的情况下学习新的信息；(3)在没有预设最优条件的情况下学习；(4)表示输入在线数据的拓扑结构；(5)学习表示每个类所需的节点数。人工数据和高维真实数据的实验表明，该方法能够实现高识别率、高速度和低内存的分类。
A new self-organizing incremental network is designed for online supervised learning. During learning of the network, an adaptive similarity threshold is used to judge if new nodes are needed when online training data are introduced into the system. Nodes caused by noise are deleted to decrease the misclassification. The proposed network, which is robust to noisy training data, suits the following tasks: (1) online or even life-long supervised learning; (2) incremental learning, i.e., learning new information without destroying old learned information; (3) learning without any predefined optimal condition; (4) representing the topology structure of inputting online data; and (5) learning the number of nodes needed to represent every class. Experiments of artificial data and high-dimension realworld data show that the proposed method can achieve classification with a high recognition ratio, high speed, and low memory.