Table 1 Top

Recall MSE - Moderate corruption
	White noise		Pixel dropout		Mask
	CIFAR-10	TinyImgNet	CIFAR-10	TinyImgNet	CIFAR-10	TinyImgNet
GPCN(offline)¹⁹	0.0121^(±0.0001)	0.0067^(±0.0004)	0.0001^(±0.0000)	0.0000^(±0.0000)	0.0009^(±0.0000)	0.0001^(±0.0000)
BayesPCN¹⁹	0.0337^(±0.0007)	0.6606^(±0.0267)	0.0001^(±0.0000)	0.0000^(±0.0000)	0.0019^(±0.0000)	0.0000^(±0.0000)
BayesPCN(forget)¹⁹	0.0188^(±0.0002)	0.0176^(±0.0001)	0.0019^(±0.0000)	0.0008^(±0.0000)	0.0465^(±0.0001)	0.0235^(±0.0001)
MHN	0.1457^(±0.0097)	0.0955^(±0.0073)	0.4512^(±0.0558)	0.4545^(±0.0456)	0.5538^(±0.0149)	0.6101^(±0.0091)
MHN-Manhtn	0.0000^(±0.0000)	0.0000^(±0.0000)	0.0000^(±0.0000)	0.0000^(±0.0000)	0.0020^(±0.0028)	0.0016^(±0.0012)
MHN-GradInf¹⁹	0.0000^(±0.0000)	0.0000^(±0.0000)	0.0000^(±0.0000)	0.0000^(±0.0000)	0.0000^(±0.0000)	0.0001^(±0.0000)
SQHN L1	0.0000^(±0.0000)	0.0000^(±0.0000)	0.0000^(±0.0000)	0.0000^(±0.0000)	0.0000^(±0.0000)	0.0000^(±0.0000)
SQHN L2	0.0002^(±0.0001)	0.0000^(±0.0000)	0.0000^(±0.0000)	0.0002^(±0.0002)	0.0001^(±0.0000)	0.0008^(±0.0002)
SQHN L3	0.1076^(±0.0022)	0.0000^(±0.0000)	0.0000^(±0.0000)	0.0000^(±0.0000)	0.0000^(±0.0000)	0.0000^(±0.0000)

Recall MSE of 1024 images from CIFAR-10 and Tiny ImageNet data sets. At test time, images are either corrupted with white noise (variance 0.2), 25% of the pixels dropped, or the right 25% of the pixels masked. Bottom Recall MSE of 128 images from CIFAR-10 and Tiny ImageNet data sets. At test time, images are either corrupted with white noise (variance 0.8), 75% of the pixels dropped, or the right 75% of the pixels masked.

Search