notesum.ai

Published at April 22

Convolutional Differentiable Logic Gate Networks

NeurIPS

Released Date: April 22, 2024

Authors: Felix Petersen¹, Hilde Kuehne², Christian Borgelt³, Julian Welzel⁴, Stefano Ermon⁵

Aff.: ¹Stanford University InftyLabs Research; ²Tuebingen AI Center MIT-IBM Watson AI Lab; ³University of Salzburg; ⁴InftyLabs Research; ⁵Stanford University

Arxiv: https://openreview.net/pdf/550935e8b4e775076ce2310d9d089be095ad0708.pdf

Method	Acc.	# Gates
DiffLogic Net (medium) [petersen2022difflogic]	57.39%	0.51 M
DiffLogic Net (largest) [petersen2022difflogic]	62.14%	5.12 M
Conv. TTNet (small) [benamira2023scalable]	50.10%	0.57 M
Conv. TTNet (large) [benamira2023scalable]	70.75%	189 M
FINN CNV [umuroglu2017finn]	80.10%	901 M
LUTNet [wang2020lutnet_tc]	84.95%	1 290 M
XNOR-Net [rastegari2016xnor] (NIN) [yu2023xnor]	86.28%	1 780 M
RebNet (1 residual) [ghasemzadeh2018rebnet]	80.59%	2 270 M
RebNet (2 residuals) [ghasemzadeh2018rebnet]	85.94%	2 830 M
BinaryNet [hubara2016binarized]	88.60%	4 090 M
Zhao et al. [zhao-bnn-fpga2017]	88.54%	4 940 M
FBNA CNV [guo2018fbna]	88.61%	5 540 M
Hirtzlin et al. [hirtzlin2019stochastic]	91. %	87 400 M
LogicTreeNet-S	60.38%	0.40 M
LogicTreeNet-M	71.01%	3.08 M
LogicTreeNet-B	80.17%	16.0 M
LogicTreeNet-L	84.99%	28.9 M
LogicTreeNet-G	86.29%	61.0 M