Table 1. Convolutional neural network architectures used in this paper.

For ResNet34, the features considered involve only layers that are not “skipped over”, while for Inception_v3 only elementary layers with non-negative outputs (ReLU, max/avg-pooling) that are not on the auxiliary branch.

Network	Number of features flz	Number of trainable parameters	Reference
AlexNet	2,816	57,163,623	Krizhevsky, 2012 [7]
VGG11	6,976	128,926,119	Simonyan, 2014 [28]
VGG13	7,360	129,110,631	Simonyan, 2014 [28]
VGG16	9,920	134,420,327	Simonyan, 2014 [28]
			(0)Conv2d(3,64) (1)ReLU (2)Conv2d(64,64) (3)ReLU (4)MaxPool2d (5)Conv2d(64,128) (6)ReLU (7)Conv2d(128,128) (8)ReLU (9)MaxPool2d (10)Conv2d(128,256) (11)ReLU (12)Conv2d (256,256) (13)ReLU (14)Conv2d(256,256) (15)ReLU (16)MaxPool2d (17)Conv2d (256,512) (18)ReLU (19)Conv2d(512,512) (20)ReLU (21)Conv2d(512,512) (22)ReLU (23)MaxPool2d (24)Conv2d(512,512) (25)ReLU (26)Conv2d(512,512) (27)ReLU (28)Conv2d(512,512) (29)ReLU (30)MaxPool2d (31)Linear(25088,4096) (32)ReLU (33)Dropout(0.5) (34)Linear(4096,4096) (35)ReLU (36)Dropout(0.5) (37)Linear(4096,39)
VGG19	12,480	139,730,023	Simonyan, 2014 [28]
VGG11_bn	9,728	128,931,623	Simonyan, 2014 [28]; Ioffe, 2015 [36]
VGG13_bn	10,304	129,116,519	Simonyan, 2014 [28]; Ioffe, 2015 [36]
VGG16_bn	14,144	134,428,775	Simonyan, 2014 [28]; Ioffe, 2015 [36]
VGG19_bn	17,984	139,741,031	Simonyan, 2014 [28]; Ioffe, 2015 [36]
ResNet34	28,992	21,304,679	He, 2016 [29]
Inception_v3	27,712	24,453,166	Szegedy, 2016 [30]
VGG16_1FC	9,920	15,693,159	This paper: same convolutional part as VGG16, but with a single fully connected layer: Conv(VGG16); Dropout(0.5); Linear(25088,39)
VGG16_avg1FC	10,432	14,734,695	This paper: the convolutional part as VGG16, followed by an average pooling layer and a single fully connected layer: Conv(VGG16); AvgPool2d(7,7,512;1,1,512); Dropout(0.5); Linear(512,39)