基于深度互学的多任务学习

肖洪湖,黄成泉,周训会,董红来,周丽华

Multi-task learning based on deep mutual learning

Honghu XIAO,Chengquan HUANG,Xunhui ZHOU,Honglai DONG,Lihua ZHOU

表 4 各方法在NYUv2数据集上的模型复杂度比较

Tab.4 Comparisons of model complexity of different methods on NYUv2 dataset

加权方案 方法 网络 从头开始训练网络 预训练网络 ttr/s Np/106 ttr/s Np/106 DWA Independent Net 1 13.83 96.69 16.34 96.69 Net 2 17.18 135.25 19.05 135.25 MDML Net 1&Net 1 24.47 193.38 22.99 193.38 MDML Net 2&Net 2 28.74 270.50 22.76 270.50 FAMO Independent Net 1 17.51 96.69 16.05 96.69 Net 2 20.81 135.25 21.45 135.25 MDML Net 1&Net 1 26.87 193.38 26.03 193.38 MDML Net 2&Net 2 32.98 270.50 35.70 270.50 OTW&MLW Independent(OKD) Net 1 38.02 200.99 36.89 200.99 Net 2 51.48 316.67 51.24 316.67 MDML Net 1&Net 1 21.33 193.38 24.94 193.38 MDML Net 2&Net 2 16.10 270.50 22.44 270.50