TABLE XXI.

Impact of NAS Vs. Handcrafted Models

Application	Dataset	Method	Accuracy	Latency/MAC	Flash
Inertial Odometry◦	OxIOD	L-IONet TCN [231]	2.82 m	13.9M	183 kB
		RoNiN TCN [232]	0.42m	220M	2.1 MB
		TinyOdom TCN [59]	1.24m-1.37 m	4.64M-8.92M	71 kB-118 kB
Audio Keyword Spotting	Speech Commands	DS-CNN [9]	92%	5.54M	52.5 kB
		MicroNets DS-CNN [8]	95.3% - 96.5%	16M-129M	102 kB-612 kB
		μNAS-CNN [122]	95.4-95.6%	1.1M	19 kB-37 kB
Image Recognition	Visual Wake Words*	MBNetv2 [8]	86%	0.46s	375 kB
		MicroNets MBNetv2 [8]	78.1%-88%	0.08s-1.13s	230 kB-833 kB
	MNIST-10*	Bonsai [57]	94.4%	8.9 mS	1.97 kB
		SpArSe-CNN [86]	95.8%-97%	27mS-286mS	2.4kB-15.9kB
	CIFAR-10B,*	Bonsai [57]	73%	8.2 mS	1.98 kB
		SpArSe-CNN [86]	70.5%-73.4%	0.49s-2.52s	2.7 kB-9.9 kB
		μNAS-CNN [122]	77.5%	-	0.69kB

^◦Accuracy metric is relative trajectory error [232] (lower is better)

^*Device: STM32 Cortex-M4 and M7

^BBinary dataset

Handcrafted models