notesum.ai

Published at December 9

DEX: Data Channel Extension for Efficient CNN Inference on Tiny AI Accelerators

cs.LG

Released Date: December 9, 2024

Authors: Taesik Gong¹, Fahim Kawsar¹, Chulhong Min²

Aff.: ¹Nokia Bell Labs; ²UNIST

Arxiv: http://arxiv.org/pdf/2412.06566v1

Refer to caption

Dataset	Method	SimpleNet	WideNet	EfficientNetV2	MobileNetV2	AVG (%)
ImageNette	Downsampling	57.8 ± 1.2	61.8 ± 0.2	51.3 ± 0.5	62.0 ± 0.7	58.2
	CoordConv	58.0 ± 1.1	61.7 ± 0.2	51.9 ± 0.1	61.6 ± 0.3	58.3
	CoordConv (r)	55.4 ± 1.5	61.4 ± 0.2	51.7 ± 1.0	61.2 ± 1.1	57.4
	DEX (ours)	61.4 ± 0.6	65.6 ± 0.6	56.8 ± 0.5	64.4 ± 0.6	62.0
Caltech101	Downsampling	54.6 ± 2.1	55.8 ± 1.2	38.6 ± 0.9	51.4 ± 1.6	50.1
	CoordConv	53.8 ± 1.6	56.5 ± 0.1	38.7 ± 0.2	49.8 ± 0.5	49.7
	CoordConv (r)	52.7 ± 0.5	56.0 ± 1.7	38.2 ± 1.0	49.7 ± 1.2	49.1
	DEX (ours)	56.9 ± 1.3	61.1 ± 1.4	45.9 ± 1.9	53.3 ± 1.7	54.3
Caltech256	Downsampling	19.8 ± 0.6	20.8 ± 0.5	14.7 ± 0.4	22.4 ± 1.0	19.4
	CoordConv	19.8 ± 0.5	21.3 ± 0.8	14.8 ± 0.8	22.7 ± 0.8	19.6
	CoordConv (r)	20.0 ± 1.6	20.9 ± 0.6	14.5 ± 0.3	22.7 ± 0.4	19.5
	DEX (ours)	22.8 ± 0.5	22.9 ± 0.9	18.3 ± 0.9	26.3 ± 0.5	22.6
Food101	Downsampling	16.0 ± 0.4	17.7 ± 0.7	12.1 ± 0.2	22.4 ± 0.6	17.1
	CoordConv	16.1 ± 0.8	17.7 ± 0.3	12.0 ± 0.1	21.7 ± 0.3	16.9
	CoordConv (r)	16.3 ± 0.4	17.3 ± 0.6	12.0 ± 0.6	20.9 ± 0.3	16.6
	DEX (ours)	18.4 ± 0.4	20.9 ± 0.4	16.4 ± 0.1	23.3 ± 1.1	19.8