机构地区: 华南理工大学计算机科学与工程学院
出 处: 《计算机科学》 2009年第8期212-214,242,共4页
摘 要: 迁移学习能够有效地在相似任务之间进行信息的共享和迁移。之前针对多任务回归的迁移学习研究大多集中在线性系统上。针对非线性回归问题,提出了一种新的多任务回归模型——HiRBF。HiRBF基于层次贝叶斯模型,采用RBF神经网络进行回归学习,假设各个任务的输出层参数服从某种共同的先验分布。根据各个任务是否共享隐藏层,在构造HiRBF模型时有两种可选方案。在实验部分,将两种方案进行了对比,也将HiRBF与两种非迁移学习算法进行了对比,实验结果表明,HiRBF的预测性能大大优于其它两个算法。 Multi-task learning utilizes labeled data from other "similar" tasks and can achieve efficient knowledge-sharing between tasks. Previous research mainly focused on multi-task learning for linear regression. A novel Bayesian multi-task learning model for non-linear regression, i. e. HiRBF, was proposed. HiRBF is constructed under a hierarchical Bayesian framework. According to whether the input-to-hidden is shared by all tasks or not,we have two options to build the HiRBF model. There is a comparison between them in the experiment section. The HiRBF algorithm is also compared with two transfer-unaware approaches. The experiments demonstrate that HiRBF significantly outperforms the others.
领 域: [自动化与计算机技术] [自动化与计算机技术]