基于距离度量损失框架的半监督学习方法

基于距离度量损失框架的半监督学习方法

刘半藤,叶赞挺,秦海龙,王柯,郑启航,王章权

Semi-supervised learning method based on distance metric loss framework

Ban-teng LIU,Zan-ting YE,Hai-long QIN,Ke WANG,Qi-hang ZHENG,Zhang-quan WANG

表 2 各方法在CIFAR-10、CIFAR-100、SVHN、STL-10数据集中的准确率

Tab.2 Accuracy of each method on CIFAR-10, CIFAR-100, SVHN, STL-10 datasets

方法	CIFAR-10				CIFAR-100				SVHN				STL-10
方法	N_l = 40	N_l = 250	N_l = 4000		N_l = 400	N_l = 1000	N_l = 10000		N_l = 40	N_l = 250	N_l = 4000		N_l = 1000	N_l = 5000
VAT	27.41	53.97	82.59		5.12	12.89	53.20		35.64	79.78	90.06		68.77	81.26
MeanTeacher	26.54	55.52	83.79		4.94	12.13	55.13		36.45	78.43	93.58		69.01	82.80
MixMatch	41.47	74.08	90.74		9.15	32.39	65.87		47.45	76.02	93.50		79.59	88.41
Remixmatch	42.32	78.71	92.20		12.98	44.56	72.33		52.55	81.92	94.67		84.33	92.89
UDA	44.55	81.66	92.45		10.66	40.72	71.18		47.37	84.31	95.54		82.34	92.74
UDA_unify	49.69	84.22	93.87		18.06	46.12	72.67		53.14	87.32	95.77		87.21	94.34
Fixmatch	50.17	86.28	93.11		12.65	42.95	72.08		56.41	88.75	95.94		89.68	94.58
Fixmatch_unify	55.24	89.94	94.74		18.51	47.79	73.38		65.86	90.09	96.02		93.44	95.25
Flexmatch	52.70	85.08	93.97		27.22	55.12	77.84		62.03	90.42	95.83		92.36	96.06
Unify Loss	61.19	90.33	94.25		35.35	61.75	78.11		67.35	92.47	96.17		93.42	96.51