Development and Applications of Learning Algorithms for Neural Networks

神经网络学习算法的开发和应用

基本信息

批准号：
02650235
负责人：
TAKAHASHI Haruhisa
金额：
$ 1.41万
依托单位：
The University of Elector-Communications
依托单位国家：
日本
项目类别：
Grant-in-Aid for General Scientific Research (C)
财政年份：
1990
资助国家：
日本
起止时间：
1990 至 1991
项目状态：
已结题

来源：
https://kaken.nii.ac.jp/grant/KAKENHI-PROJECT-02650235/
关键词：
Neural network Learning Algorithm Back Propagation Recurrent Propagation 線型分離線型分離可能性

项目摘要

(1)It is mathematically investigated as to what kind of internal representations are separable. by a single output unit of a three layer feed next neural network. A topologically described necessary and sufficient condition is shown for partitions of input spaces to be classified by the output unit. Then an efficient algorithm is proposed for checking if a given partition of the input space is resulted in linear separation at the output unit.(2)We propose in this paper a new recurrent propagation learning algorithm. A biologically plausible neurodynamics is derived from which a quick algorithm to compute fixed points is obtained applying an approximation of stochastic process. The sensitivity of the networks is also obtained and from that a new recurrent propagation teaming algorithm is proposed. Our algorithm can run 10 times more quick over the previous ones. Furthermore, some unstable problems in recurrent propagation are overcome. Some simulation results are given to compare the re … More current propagations with Backpropagation and Mean Field Networks.(3)Associative memory is realized by the, recurrent propagation Teaming rule as equilibrium states ; of the network. Since the number of hidden units can be increased unrestrictedly, the information capacity can be sufficiently large. This is the main advantage of the proposed network compared with previous ones that have no hidden units. Some printer fonts are used for patterns to be memorized in our experiments. The data show that the more the number of hidden units, the more the rate of memorization and correct association.(4)A new teaming network is proposed in this chapter. Although the Backpropagation Teaming method has been succeeded in many applications, some difficulties exists, such as the local minimal problem, too long Teaming Teaming times, and capability of approximation to continuous mappings. We combine the competitive learning and supervised teaming methods in order to approximate continuous mappings. Some simulation results show that the proposed teaming algorithm works extremely more quick than the Back Propagation method and has reliable convergence property. Less

（1）在数学上研究了哪种内部表示形式是分开的。由三层馈送下一个神经网络的单个输出单元。为要由输出单元分类的输入空间的分区显示了拓扑描述的必要条件。然后提出了一种有效的算法，用于检查输入空间的给定分区是否导致输出单元的线性分离。（2）我们在本文中提出了一种新的经过重复的传播学习算法。从生物学上合理的神经动力学得出了该算法来计算固定点的快速算法，以应用随机过程的近似值。还获得了网络的敏感性，并提出了一种新的经常性传播组合算法。我们的算法比以前的算法可以更快地运行10倍。此外，复发传播中的一些不稳定问题正在克服。给出了一些仿真结果，以比较RE…（3）相关记忆是由相当的繁殖统一规则实现的，作为等效状态；网络。由于可以不受限制地增加隐藏单元的数量，因此信息容量可能足够大。与没有隐藏单位的网络相比，这是拟议网络的主要优点。一些打印机字体用于在我们的实验中记住的模式。数据表明，隐藏单元的数量越多，内存率越多。（4）本章提出了一个新的团队网络。尽管反向传播的组合方法在许多应用中都取得了成功，但存在一些困难，例如本地最小问题，太长的团队组合时间以及近似能力连续映射的能力。我们结合了竞争性学习和监督的团队方法，以近似连续映射。一些仿真结果表明，所提出的团队算法比背部传播方法更快地工作，并且具有可靠的收敛属性。较少的