迁移学习能以相关领域中的标注数据为基础,提升目标领域的学习效果。当领域间的数据分布差异很大时,会导致严重的负迁移问题。如何充分捕获源域和目标域之间的相似性,进一步挖掘更多有效信息,最终提高目标域的预测精度,是一个值得探索...迁移学习能以相关领域中的标注数据为基础,提升目标领域的学习效果。当领域间的数据分布差异很大时,会导致严重的负迁移问题。如何充分捕获源域和目标域之间的相似性,进一步挖掘更多有效信息,最终提高目标域的预测精度,是一个值得探索的问题。该文从细粒度主动迁移的视角,提出一种深度子领域迁移学习(Deep subdomain transfer learning,DSTL)算法,能迭代优化源域和目标域之间的相似性,提升模型预测性能。该文首先提出一种伪标签生成策略,对所有样本进行子领域的划分;制定中心+边缘的主动查询策略,获得关键代表性实例的真实标签;设计一种迭代分布优化策略,实现源域和目标域的子领域对齐,避免负迁移。将DSTL算法与传统迁移学习算法以及当前最新的深度迁移学习算法在主流的基准数据集上进行了测试。统计分析的结果表明,该文所提算法能实现性能的有效提升,扩大模型在实际应用中的适用范围。展开更多
Traditionally,plant distribution is thought to be closely related to environmental factors.But recently,it is found that Populus,quite different from other plant taxa,adapted to negative environmental changes,and succ...Traditionally,plant distribution is thought to be closely related to environmental factors.But recently,it is found that Populus,quite different from other plant taxa,adapted to negative environmental changes,and successfully migrated to different climate zones from its origin places of warm temperate zone.Conversely,Metasequoia is gradually tending to extinction from the Miocene to Quaternary.Based on above contrary cases,two response patterns of plant to negative environmental changes are proposed.One is active adaptation represented by Populus,the other is passive adaptation represented by Metasequoia.The plants of passive strategy characterized for desert prevention might be easily replaced by those of active strategy characterized for desert utilization.Fast growing plants,such as Populus with characteristics of drought and salt tolerance,wind and sand resistance,are selected in Tarim Basin in southern Xinjiang,China,as a good example of desert utilization in the construction of new highways and towns,not only serve as farmland shelterbelt in sandy area.In addition,Populus with high-altitude and cold adaptation has also been selected as an ideal tree planted in Tibet.Therefore,the idea of using Populus as one of the preferred pioneer trees to colonize Mars is proposed.展开更多
文摘迁移学习能以相关领域中的标注数据为基础,提升目标领域的学习效果。当领域间的数据分布差异很大时,会导致严重的负迁移问题。如何充分捕获源域和目标域之间的相似性,进一步挖掘更多有效信息,最终提高目标域的预测精度,是一个值得探索的问题。该文从细粒度主动迁移的视角,提出一种深度子领域迁移学习(Deep subdomain transfer learning,DSTL)算法,能迭代优化源域和目标域之间的相似性,提升模型预测性能。该文首先提出一种伪标签生成策略,对所有样本进行子领域的划分;制定中心+边缘的主动查询策略,获得关键代表性实例的真实标签;设计一种迭代分布优化策略,实现源域和目标域的子领域对齐,避免负迁移。将DSTL算法与传统迁移学习算法以及当前最新的深度迁移学习算法在主流的基准数据集上进行了测试。统计分析的结果表明,该文所提算法能实现性能的有效提升,扩大模型在实际应用中的适用范围。
基金Supported by projects of the Open Research Fund of State Key Laboratory of Modern Paleontology and Stratigraphy,Nanjing Institute of Geology and Palaeontology,CAS(No.213127)the National Natural Science Foundation of China(No.31470324)Higher Education Teaching Reform Project of Shenyang Normal University,2014(No.052/51400301).
文摘Traditionally,plant distribution is thought to be closely related to environmental factors.But recently,it is found that Populus,quite different from other plant taxa,adapted to negative environmental changes,and successfully migrated to different climate zones from its origin places of warm temperate zone.Conversely,Metasequoia is gradually tending to extinction from the Miocene to Quaternary.Based on above contrary cases,two response patterns of plant to negative environmental changes are proposed.One is active adaptation represented by Populus,the other is passive adaptation represented by Metasequoia.The plants of passive strategy characterized for desert prevention might be easily replaced by those of active strategy characterized for desert utilization.Fast growing plants,such as Populus with characteristics of drought and salt tolerance,wind and sand resistance,are selected in Tarim Basin in southern Xinjiang,China,as a good example of desert utilization in the construction of new highways and towns,not only serve as farmland shelterbelt in sandy area.In addition,Populus with high-altitude and cold adaptation has also been selected as an ideal tree planted in Tibet.Therefore,the idea of using Populus as one of the preferred pioneer trees to colonize Mars is proposed.
