Abdulnabi Abrar H. Wang Gang Lu Jiwen Jia Kui

Abstract

This paper proposes a joint multi-task learning algorithm to better predictattributes in images using deep convolutional neural networks (CNN). Weconsider learning binary semantic attributes through a multi-task CNN model,where each CNN will predict one binary attribute. The multi-task learningallows CNN models to simultaneously share visual knowledge among differentattribute categories. Each CNN will generate attribute-specific featurerepresentations, and then we apply multi-task learning on the features topredict their attributes. In our multi-task framework, we propose a method todecompose the overall model's parameters into a latent task matrix andcombination matrix. Furthermore, under-sampled classifiers can leverage sharedstatistics from other classifiers to improve their performance. Naturalgrouping of attributes is applied such that attributes in the same group areencouraged to share more knowledge. Meanwhile, attributes in different groupswill generally compete with each other, and consequently share less knowledge.We show the effectiveness of our method on two popular attribute datasets.