Predicting Multiple Attributes via Relative Multi-task Learning

被引:115
作者
Chen, Lin [1 ]
Zhang, Qiang [1 ]
Li, Baoxin [1 ]
机构
[1] Arizona State Univ, Tempe, AZ 85281 USA
来源
2014 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR) | 2014年
关键词
D O I
10.1109/CVPR.2014.135
中图分类号
TP18 [人工智能理论];
学科分类号
081104 ; 0812 ; 0835 ; 1405 ;
摘要
Relative attributes learning aims to learn ranking functions describing the relative strength of attributes. Most of current learning approaches learn ranking functions for each attribute independently without considering possible intrinsic relatedness among the attributes. For a problem involving multiple attributes, it is reasonable to assume that utilizing such relatedness among the attributes would benefit learning, especially when the number of labeled training pairs are very limited. In this paper, we proposed a relative multi-attribute learning framework that integrates relative attributes into a multi-task learning scheme. The formulation allows us to exploit the advantages of the state-of-the-art regularization-based multi-task learning for improved attribute learning. In particular, using joint feature learning as the case studies, we evaluated our framework with both synthetic data and two real datasets. Experimental results suggest that the proposed framework has clear performance gain in ranking accuracy and zero-shot learning accuracy over existing methods of independent relative attributes learning and multi-task learning.
引用
收藏
页码:1027 / 1034
页数:8
相关论文
共 25 条
[1]  
Ando Rie Kubota, 2005, J MACH LEARN RES
[2]  
[Anonymous], 2011, Advances in neural information processing systems
[3]  
[Anonymous], 2010, Graph-structured multi-task regression and an efficient optimization method for general fused lasso
[4]  
[Anonymous], INT J COMPUTER VISIO
[5]  
Bach F. R., CLUSTERED MULTITASK
[6]  
Bakker B., 2003, J MACH LEARN RES DEC
[7]  
Baxter D. A., 2007, SIMULATOR NEURAL NET
[8]  
Berg T. L., 2010, PROC
[9]  
Caruana Rich, 1997, Multitask Learning *
[10]  
Chen J., 2011, P KDD 11