Convolutional Neural Network-Based Discriminator for Outlier Detection

Table 12

Average test accuracy on MNIST over the last ten epochs. The top part of the table presents the results that were adapted as published in [29]. The bottom part shows the results achieved by our implementation.

Method	Flipping rate
	Symmetry-20%	Symmetry-50%	Pair-45%
Standard	94.05 (±0.16)	66.05 (±0.61)	56.52 (±0.55)
Bootstrap	94.40 (±0.26)	67.55 (±0.53)	57.23 (±0.73)
S-model	98.31 (±0.11)	62.29 (±0.46)	56.88 (±0.32)
F-correction	98.80 (±0.12)	79.61 (±1.96)	0.24 (±0.03)
Decoupling	95.70 (±0.02)	81.15 (±0.03)	58.03 (±0.07)
MentorNet	96.70 (±0.22)	90.05 (±0.30)	80.88 (±4.45)
Coteaching	97.25 (±0.03)	91.32 (±0.06)	87.63 (±0.21)
EBF	98.75 (±0.29)	98.27 (±0.39)	88.91 (±0.62)
Discriminator (2S + DA)	96.27 (±0.20)	93.07 (±0.73)	93.29 (±0.32)
Discriminator (5S)	89.12 (±0.23)	87.67 (±0.51)	93.26 (±0.20)
Discriminator (5S + DA)	96.98 (±0.34)	96.33 (±0.36)	97.13 (±0.18)