Convolutional Neural Network-Based Discriminator for Outlier Detection

Table 15

Effect of the discriminator model on the training time.

		20% symm. noise			50% symm. noise			45% pair noise
		Test acc.	Time (sec)	Epochs	Test acc.	Time (sec)	Epochs	Test acc.	Time (sec)	Epochs
MNIST	Standard	93.15	15.1 K	148	64.41	29.8 K	300	54.39	19.5 K	197
	Discriminator	95.60	273.43	4	94.83	523.97	12	96.02	198.38	4
CIFAR-10	Standard	71.38	8.2 K	90	47.50	15.2 K	158	47.58	8.6 K	89
	Discriminator	83.06	5.1 K	92	78.69	5.4 K	139	80.00	5.3 K	129
CIFAR-100	Standard	37.78	27.3 K	300	16.60	27.7 K	300	27.63	27.7 K	300
	Discriminator	50.78	10.8 K	300	42.46	7.6 K	300	44.69	8.2 K	300
Fashion	Standard	85.76	14.7 K	145	56.93	30.3 K	300	54.67	19.3 K	190
	Discriminator	84.34	1.1 K	19	79.58	1.1 K	29	85.09	793.46	18
Traffic Signs	Standard	88.55	21.3 K	300	73.64	21.4 K	300	51.44	21.2 K	300
	Discriminator	92.95	1.3 K	21	92.12	1.7 K	51	91.96	3.7 K	101

K refers to a thousand of seconds.