分布：Ran(−1, 1)

Cifar-100

Mini-ImageNet

模型

VoVNet-57 [21]

ResNeSt50 [23]

VoVNet-57 [21]

ResNeSt50 [23]

现有方法

64.05%

66.73%

72.56%

69.87%

DLC-UG的方法

ε = (−0.0, 0.0)

72.68%

74.69%

79.38%

77.67%

ε = (−0.2, 0.2)

67.36%

70.10%

75.78%

73.15%

ε = (−0.4, 0.4)

66.04%

68.49%

74.42%

70.38%

DLC-UG-opt的方法

ε = (−0.0, 0.0)

73.08%

74.99%

79.58%

77.97%

ε = (−0.2, 0.2)

72.08%

74.23%

78.78%

76.24%

ε = (−0.4, 0.4)

71.72%

72.12%

78.07%

74.97%