文章摘要
基于降噪自动编码器的多任务优化算法--2018CPCC会议推荐论文
Multi-task Optimization through Denoising Autoencoding
投稿时间:2018-10-17  修订日期:2018-11-27
DOI:
中文关键词: 多任务优化  多任务学习  降噪自动编码器  单任务优化  基于种群的搜索算法
英文关键词: Multi-task Optimization  Denoising Autoencoder  Single-task Optimization  Multi-task Learning  Population-based Search Algorithm
基金项目:国家自然科学青年基金(61603064);中央高校前沿交叉项目(106112017CDJQJ188828);重庆市前沿基础应用面上项目(cstc2017jcyjAX0319).
作者单位E-mail
尚青霞 重庆大学计算机学院 974163837@qq.com 
周磊 重庆大学计算机学院  
冯亮 重庆大学计算机学院 liangf@cqu.edu.cn 
摘要点击次数: 74
全文下载次数: 
中文摘要:
      人类的学习行为通常可以多个任务同时进行,且一个学习任务的知识往往能加速另一个相似任务的学习,受此启发,多任务学习(MTL)已广泛地被国内外学者研究。与多任务学习的动机类似,多任务优化(MTO)是在传统基于种群的优化算法的基础上提出的一种新型优化方法,该方法的目的是同时在线执行多个任务,从一个任务中获取知识,以帮助另一个任务,进行任务间知识迁移,进而提高多任务的优化性能。本文基于降噪自动编码器提出了一种新的MTO算法,该算法从降噪自动编码器中推演出一种具有闭式解的降噪自动编码器,并利用此编码器显式地对多任务构建任务映射,从而使新提出的MTO算法能够利用不同的基于种群的优化算法的搜索偏好,实现高效的多任务优化性能。本文采用常用的MTO基准进行综合性实验,实验验证了新提出的MTO算法的有效性。
英文摘要:
      Inspired by the remarkable ability of human learning which is able to perform multiple tasks simultaneously, and apply the knowledge gained from one task to help another, multi-task learning (MTL) has been proposed and well-studied in the literature. With similar motivation as MTL, multi-task optimization (MTO) has recently been proposed as a new paradigm for optimization. In contrast to the traditional single-task optimization paradigm, MTO conducts the optimization process on multiple problems simultaneously. It aims to improve the optimization performance across multiple problems by seamlessly transferring knowledge between them online. In this paper, we present a new MTO algorithm with denoising autoencoding. In particular, a denoising autoencoder is derived for building mappings across tasks explicitly, thus making the proposed MTO algorithm be able to leverage the diverse biases induced by different optimization solvers. To evaluate the performance of the proposed algorithm, comprehensive empirical studies on complex MTO benchmark sets have been presented.
View Fulltext   查看/发表评论  下载PDF阅读器
关闭