Multi-target deep neural networks: Theoretical analysis and implementation

NEUROCOMPUTING

卷:273

Article

Zeng, Zeng¹  Liang, Nanying¹  Yang, Xulei²  Hoi, Steven³ 

[1] Inst Infocomm Res, 1 Fusionopolis Way,21-01 Connexis South Tower, Singapore 138632, Singapore

[2] Inst High Performance Comp, 1 Fusionopolis Way,16-16 Connexis, Singapore 138632, Singapore

[3] Singapore Management Univ, SIS, 81 Victoria St, Singapore 188065, Singapore

关键词: Deep neural networks;    Multi-target deep learning;    Object detection;    Segmentation;    Learning path;   

DOI  :  10.1016/j.neucom.2017.08.044

来源: Elsevier

【 摘 要 】

In this work, we propose a novel deep neural network referred to as Multi-Target Deep Neural Network (MT-DNN). We theoretically prove that different stable target models with shared learning paths are stable and can achieve optimal solutions respectively. Based on GoogleNet, we design a single model with three different targets, one for classification, one for regression, and one for masks that is composed of 256 x 256 sub-models. Unlike bounding boxes used in ImageNet, our single model can draw the shapes of target objects, and in the meanwhile, classify the objects and calculate their sizes. We validate our single MT-DNN model via rigorous experiments and prove that the multiple targets can boost each other to achieve optimization solutions. (C) 2017 Elsevier B.V. All rights reserved.

【 授权许可】

Free