【计算机科学】【2005.12】神经网络在生物数据中的应用

本文为英国伦敦大学（作者：Aristoklis.D. Anastasiadis）的博士论文，共184页。

在分类问题中训练神经网络，特别是当涉及生物数据时，是一项非常具有挑战性的任务。到目前为止，已经提出了许多训练算法来改善神经网络的性能。一种流行的方法是使用批量学习，对每个权值使用不同的自适应学习率。大多数现有的该类算法都是基于启发式的，并且它们不能保证从任何初始权值集收敛到局部最小值。此外，它们经常会收敛到局部极小值，特别是当开始训练的初值远离最小值时。为了改善这种情况，本博士学位论文提出了克服这些问题的新方法。

本文提出了一类基于复合非线性雅可比过程的自适应学习率符号化方案；提出了一种自适应策略，保证在搜索方向为下降方向时批量误差的降低。此外，它使新算法具有全局收敛性质，即从任意远处的出发点收敛到局部最小值。然后，在全局搜索方法的背景下，通过提出一种结合确定性搜索、随机搜索以及自适应学习速率的混合学习方案，来处理偶尔收敛到局部极小值的问题。在非扩展统计力学的背景下探索随机搜索方法，通过使用q-非广延熵指数产生的扰动修正训练过程中的误差表面。所提出的算法用于训练生物和生物信息学数据集中的前馈神经网络及多种神经网络集合。

Training neural networks in classiflcation problems, especially whenbiological data are involved, is a very challenging task. Many trainingalgorithms have been proposed so far to improve the performance of neuralnetworks. A popular approach is to use batch learning that employs a difierentadaptive learning rate for each weight. Most of the existing algorithms of thisclass are based on the use of heuristics, and they don’t guarantee convergenceto a local minimiser from any initial weights set. In addition, they convergefrequently to local minima, particularly when training starts far away from aminimiser. To alleviate this situation, this PhD thesis proposes new methodsthat overcome these problems. It proposes a new class of sign{based schemeswith adaptive learning rates that are based on the composite nonlinear Jacobiprocess. It develops an adaptation strategy that ensures the search directionis a descent one and the decrease of the batch error is guaranteed. Moreover, itequips the new algorithms with the global convergence property; i.e. it provesconvergence to a local minimiser from any remote starting point. The problem ofoccasional convergence to local minima is then dealt within the context ofglobal search methods by proposing a hybrid-learning scheme that combinesdeterministic and stochastic search, and adaptive learning rates. Stochasticsearch is explored in the context of Nonextensive Statistical Mechanics, bymodifying the error surface during training using perturbations generated bythe q-nonextensive entropic index. The proposed algorithmsare applied to train feed-forward neural networks and diverse neural ensemblesin biological and bioinformatics datasets.

1 引言
2 优化环境下的监督训练
3 基于梯度下降的训练方案
4 全局收敛训练算法
5 非扩展混合学习方案
6 生物信息学中神经网络集的训练
7 结论与未来工作展望
附录 问题描述/数据集/评估方法

下载英文原文地址：

http://page5.dfpan.com/fs/alcj9221d291f6645b2/

更多精彩文章请关注微信号：