Table 1.

Comparisons of the three candidate networks for transfer learning in terms of trainable parameter number, computational requirements for a single inference, and single-crop top 1 accuracy on the ImageNet validation dataset

	No. of trainable parameters	No. of operations needed for a single inference	Single-crop top 1 validation accuracy
GoogleNet [14]	˜̃5M (1×)	˜̃3 G-ops (1×)	˜̃68.00%
AlexNet [13]	˜̃60M (12×)	˜̃2.5 G-ops (0.83×)	˜̃54.50%
VGG-16 [15]	˜̃140M (28×)	˜̃32 G-ops (10.6×)	˜̃70.60%

Numbers from a comparative study conducted by Canziani et al. [24]