. Author manuscript; available in PMC: 2023 Nov 15.

Published in final edited form as: IEEE Sens J. 2022 Oct 5;22(22):21362–21390. doi: 10.1109/jsen.2022.3210773

TABLE XXII.

Impact of Runtime Optimizations vs. No Optimizations

Application	Dataset	Method	Accuracy	Latency/MAC	Flash
Human Activity Recognition	Custom^*	CNN-TFLM [189]	85%	58 mS	275 kB
Human Activity Recognition	Custom^*	CNN-Cube.AI [189]	85%	14 mS	192 kB
Audio Keyword Spotting	Speech Commands^*	CNN-TFLM [189]	-	380 mS	288 kB
Audio Keyword Spotting	Speech Commands^*	CNN-Cube.AI [189]	-	373 mS	247 kB
Image Recognition	ImageNet	MCUNet MbNetv2 [121]	60.3%-68.5%	68M-126M	1MB-2MB
	ImageNet	MCUNetV2 MbNetv2 [121]	64.9%-71.8%	119M-256M	1MB-2MB
	Pascal VOC	MbNetv2+CMSIS [121]	mAP: 31.6%	34M	OOS
		MCUNetV MbNetv2 [121]	mAP: 51.4%	168M	<2 MB
		MCUNetV2 MbNetv2 [121]	mAP: 64.6%	172M	<1 MB
	CIFAR-10^@	CNN [164]	80.3%	456 mS	< 1 MB
	CIFAR-10^@	CNN-CMSIS [164]	80.3%	99 mS	< 1 MB

Device: STM32 Cortex-M4

Device: STM32 Cortex-M7, OOS: Overflowed SRAM

Inline graphic Superior optimization techniques than comparing method in the same dataset class