notesum.ai

Published at October 31

ARQ: A Mixed-Precision Quantization Framework for Accurate and Certifiably Robust DNNs

cs.LG

cs.CR

cs.CV

Released Date: October 31, 2024

Authors: Yuchen Yang¹, Shubham Ugare¹, Yifan Zhao¹, Gagandeep Singh¹, Sasa Misailovic¹

Aff.: ¹University of Illinois Urbana-Champaign

Arxiv: http://arxiv.org/abs/2410.24214v1

Refer to caption

	Method	$\sigma=0.25$			$\sigma=0.50$			$\sigma=1.00$
	Method	ACR	Acc	GBitOPs	ACR	Acc	GBitOPs	ACR	Acc	GBitOPs
ResNet-50	FP 32	0.488	69.4	4244.3	0.743	62.4	4244.3	0.914	45.4	4244.3
\cdashline2-11	ARQ	0.472	70.8	63.32	0.724	61.2	63.50	0.916	46.0	63.50
\cdashline2-11	LIMPQ	0.458	68.6	63.55	0.700	58.2	63.55	0.871	44.4	63.55
	HAQ	0.460	69.0	63.56	0.715	60.8	64.15	0.880	44.8	63.55
	PACT	0.460	69.0	63.56	0.715	61.0	63.56	0.884	45.4	63.56
MobileNet-V2	FP 32	0.457	67.0	308.24	0.668	57.0	308.24	0.846	44.4	308.24
\cdashline2-11	ARQ	0.385	62.0	4.60	0.576	54.0	4.60	0.774	41.6	4.60
\cdashline2-11	LIMPQ	0.347	58.4	4.62	0.540	50.6	4.62	0.703	40.2	4.62
	HAQ	0.341	56.0	4.60	0.573	53.4	4.60	0.683	39.8	4.61
	PACT	0.376	60.2	4.62	0.564	52.4	4.62	0.757	40.2	4.62