Table 2.

The summary of DCNN architectures investigated in this work.

DCNN Model	# of Layers (#Conv. + #FC)	# of Parameters (×106)	Top-5 accuracy on ImageNet-2012	Latency [ms]	Energy/Inference [mWh]
VGG-16	13 + 3	138	90.4	98.63	0.32
ResNet-50	49 + 1	25.6	94.7	40.18	0.14
SqueezeNet	18 + 0	1.23	80.3	13.26	0.03

The specification of the DCNN models are from their respective original papers^28,29,32 except the inference latency and energy consumption. For consistency, the inference latency and energy consumption are all measured in the Nvidia Jetson TX2 embedded device.