Table 4. Carbon footprints and performance for different KD approaches.

Here, for image recognition, we use ResNet18 (teacher) and MobileNetV2 (student) models. For object detection, we use VGG16 (teacher) and MobileNetV2 (student). We report average results after running the same program five times. ↑ (↓) means higher (lower) is better.

		CIFAR 10				CIFAR 10
Loss	Method	Accuracy (%) ↑	GFLOPs (M) ↓	Energy (kWh) ↓	CO2_eq (g) ↓	Accuracy (%) ↑	GFLOPs (M) ↓	Energy (kWh) ↓	CO2_eq (g) ↓
HKD	[26]	91.78 ± 0.14	128.30	3.99	1173.44	69.70 ± 0.12	128.40	3.80	1170.40
	Ours	91.67 ± 0.23	6.42	0.21	61.76	69.40 ± 0.37	6.42	0.20	61.60
CMC	[44]	92.35 ± 0.24	128.30	3.93	1192.81	73.56 ± 0.17	128.40	3.92	1164.7
	Ours	91.65 ± 0.13	6.42	0.21	61.17	72.23 ± 0.24	6.42	0.21	61.30
DML	[80]	91.00 ± 0.04	128.30	4.01	1161.00	72.90 ± 0.04	128.40	3.74	1109.40
	Ours	91.05 ± 0.13	6.42	0.20	58.05	72.56 ± 0.15	6.42	0.19	58.04
		Tiny ImageNet				PASCAL VOC
Loss	Method	Accuracy (%) ↑	GFLOPs (M) ↓	Energy (kWh) ↓	CO2_eq (g) ↓	mAP (%) ↑	GFLOPs (M) ↓	Energy (kWh) ↓	CO2_eq (g) ↓
HKD	[26]	60.46 ± 0.07	975	19.09	5595.31	65.45 ± 0.07	1371	31.25	9235.39
	Ours	60.53 ± 0.31	48.75	1.04	304.88	65.48 ± 0.15	72.17	1.69	499.21
CMC	[44]	62.76 ± 0.17	975	20.21	5949.93	64.87 ± 0.13	1371	32.27	9527.55
	Ours	59.84 ± 0.11	48.75	1.00	293.10	64.67 ± 0.12	72.17	1.70	501.45
DML	[80]	63.30 ± 0.11	975	19.38	5748.64	65.10 ± 0.09	1371	30.24	8903.52
	Ours	63.35 ± 0.07	48.75	1.02	302.56	64.98 ± 0.24	72.17	1.68	494.64