拉普拉斯多层极速学习机

发布时间：2018-05-31 16:51

本文选题：极速学习机 + 半监督学习　；参考：《软件学报》2017年10期

【摘要】：极速学习机不仅仅是有效的分类器,还能应用到半监督学习中.但是,半监督极速学习机和拉普拉斯光滑孪生支持向量机一样,是一种浅层学习算法.深度学习实现了复杂函数的逼近并缓解了以前多层神经网络算法的局部最小性问题,目前在机器学习领域中引起了广泛的关注.多层极速学习机(ML-ELM)是根据深度学习和极速学习机的思想提出的算法,通过堆叠极速学习机-自动编码器算法(ELM-AE)构建多层神经网络模型,不仅实现了复杂函数的逼近,并且训练过程中无需迭代,学习效率高.把流形正则化框架引入ML-ELM中,提出拉普拉斯多层极速学习机算法(Lap-ML-ELM).然而,ELM-AE不能很好地解决过拟合问题.针对这一问题,把权值不确定引入ELM-AE中,提出权值不确定极速学习机-自动编码器算法(WU-ELM-AE),可学习到更为鲁棒的特征.最后,在前面两种算法的基础上提出权值不确定拉普拉斯多层极速学习机算法(WUL-ML-ELM),它堆叠WU-ELM-AE构建深度模型,并用流形正则化框架求取输出权值.该算法在分类精度上有明显提高并且不需花费太多的时间.实验结果表明,Lap-ML-ELM与WUL-ML-ELM都是有效的半监督学习算法.
[Abstract]:Extreme learning machine is not only an effective classifier, but also can be applied to semi-supervised learning. However, semi-supervised extreme learning machine is a shallow learning algorithm like Laplacian smooth twin support vector machine. Depth learning realizes the approximation of complex functions and alleviates the problem of local minimization of previous multilayer neural network algorithms, which has attracted wide attention in the field of machine learning. ML-ELM) is an algorithm based on the idea of deep learning and extreme learning machine. The multilayer neural network model is constructed by stacking pole learning machine and automatic encoder algorithm. And the training process does not need iteration, learning efficiency is high. The manifold regularization framework is introduced into ML-ELM, and a Laplacian multilayer extreme learning machine algorithm is proposed. However, ELM-AE can not solve the problem of fitting well. In order to solve this problem, the weight uncertainty is introduced into ELM-AE, and the WU-ELM-AEE algorithm is proposed, which can learn more robust features. Finally, on the basis of the above two algorithms, a Laplace multilayer learning machine with uncertain weights is proposed, which stack WU-ELM-AE to construct the depth model, and use manifold regularization framework to obtain the output weights. The algorithm improves the classification accuracy obviously and does not take too much time. The experimental results show that both Lap-ML-ELM and WUL-ML-ELM are effective semi-supervised learning algorithms.
【作者单位】：中国矿业大学计算机科学与技术学院;中国科学院计算技术研究所智能信息处理重点实验室;
【基金】：国家自然科学基金(61672522,61379101) 国家重点基础研究发展计划(973)(2013CB329502)~~
【分类号】：TP181

【相似文献】