notesum.ai

Published at November 26

Improving Resistance to Noisy Label Fitting by Reweighting Gradient in SAM

cs.LG

Released Date: November 26, 2024

Authors: Hoang-Chau Luong¹, Thuc Nguyen-Quang¹, Minh-Triet Tran¹

Aff.: ¹Software Engineering Lab, University of Science, VNUHCM

Arxiv: http://arxiv.org/abs/2411.17132v1

Refer to caption

Type	Noise	CIFAR-10		CIFAR-100
Type	Noise	SAM	SANER	SAM	SANER
Symm.	25%	$93.05_{\pm 0.17}$	$\textbf{94.08}_{\pm 0.11}$ ( $\uparrow$ 1.03)	$69.68_{\pm 0.07}$	$\textbf{72.90}_{\pm 0.21}$ ( $\uparrow$ 3.22)
	50%	$88.82_{\pm 0.08}$	$\textbf{90.60}_{\pm 0.36}$ ( $\uparrow$ 1.78)	$61.17_{\pm 0.14}$	$\textbf{66.34}_{\pm 0.11}$ ( $\uparrow$ 5.17)
Asym.	25%	$94.75_{\pm 0.28}$	$\textbf{94.83}_{\pm 0.14}$ ( $\uparrow$ 0.08)	$71.57_{\pm 0.30}$	$\textbf{74.64}_{\pm 0.13}$ ( $\uparrow$ 3.07)
	50%	$81.94_{\pm 0.71}$	$\textbf{82.25}_{\pm 1.43}$ ( $\uparrow$ 0.31)	$39.11_{\pm 0.50}$	$\textbf{40.05}_{\pm 0.51}$ ( $\uparrow$ 0.94)
Depen.	25%	$92.84_{\pm 0.18}$	$\textbf{93.67}_{\pm 0.30}$ ( $\uparrow$ 0.83)	$69.46_{\pm 0.24}$	$\textbf{72.93}_{\pm 0.29}$ ( $\uparrow$ 3.47)
	50%	$87.32_{\pm 1.17}$	$\textbf{90.01}_{\pm 0.62}$ ( $\uparrow$ 2.69)	$58.71_{\pm 0.69}$	$\textbf{66.72}_{\pm 0.75}$ ( $\uparrow$ 8.01)
Real	-	$86.33_{\pm 0.07}$	$\textbf{87.89}_{\pm 0.12}$ ( $\uparrow$ 1.56)	$62.74_{\pm 0.59}$	$\textbf{64.75}_{\pm 0.30}$ ( $\uparrow$ 2.01)