Rebuttal ICML 2025

Lowest CSL — Fig 1: Lowest CSL Scores on CIFAR-100 captures easy-to-learn, Typical (likely generalized) examples. Visualized for the first 10 classes.

Highest CSL — Fig 2: Highest CSL Scores on CIFAR-100 capture hard-to-learn, Atypical (likely memorized) examples. Visualized for the first 10 classes.

Lowest CSL ImgNet — Fig 3: Lowest CSL Scores on ImageNet captures easy-to-learn, Typical (likely generalized) examples. Visualized for the first 10 classes.

Highest CSL ImgNet — Fig 4: Highest CSL Scores on ImageNet capture hard-to-learn, Atypical (likely memorized) examples. Visualized for the first 10 classes.

Lowest CSL — Fig 5: CSL Histogram on CIFAR100, captures long tail

Highest CSL — Fig 6: CSL Histogram on ImageNet, captures long tail

**Table 1:** CSL correlation and similarity with memorization compared to other methods across CIFAR-100 and ImageNet datasets. Curv. on ImageNet RN50 was to expensive to compute within the deadline, will be included in the revised version
Dataset	Arch	Subset	Method	CS	Pearson Corr.
CIFAR-100
	Inception, Feldman & Zhang	Top 5k	Final Sample Loss	0.33	0.06
			Curv	0.87	0.16
			Loss Sensitivity	0.97	0.39
			Forget Freq.	0.96	0.29
			CSL	0.93	0.40
		All	Final Sample Loss	0.24	0.17
			Curv	0.69	0.49
			Loss Sensitivity	0.81	0.76
			Forget Freq.	0.76	0.59
			CSL	0.87	0.79
ImageNet	RN18	Top 50k	Final Sample Loss	0.70	0.09
			Curv	0.87	0.07
			Loss Sensitivity	0.96	0.14
			CSL	0.92	0.15
		All	Final Sample Loss	0.64	0.46
			Curv	0.72	0.49
			Loss Sensitivity	0.71	0.46
			CSL	0.71	0.56

	RN50, Feldman & Zhang	Top 50k	Final Sample Loss	0.78	0.12
			Curv	-	-
			Loss Sensitivity	0.79	0.04
			Forget Freq.	0.68	0.15
			CSL	0.94	0.21
		All	Final Sample Loss	0.64	0.50
			Curv	-	-
			Loss Sensitivity	0.49	0.17
			Forget Freq.	0.49	0.04
			CSL	0.79	0.64

**Table 2:** Evaluating the performance of mislabeled detection of the proposed framework against existing methods on CIFAR-10 and CIFAR-100 datasets under various levels of label noise.
Dataset	Method	Noise
Dataset	Method	15% Noise	20% Noise	25% Noise	30% Noise
CIFAR-10	Thr. Learning Time (LT)	0.4969 ± 0.0041	0.4981 ± 0.0004	0.4977 ± 0.0020	0.4988 ± 0.0028
	In Conf. [Carlini et al.]	0.6224 ± 0.0130	0.5978 ± 0.0131	0.5800 ± 0.0051	0.5669 ± 0.0106
	CL [Northcutt et al.]	0.7345 ± 0.1672	0.7169 ± 0.1539	0.6960 ± 0.1387	0.6794 ± 0.1264
	SSFT [Maini et al.]	0.9233 ± 0.0029	0.9077 ± 0.0023	0.8910 ± 0.0050	0.8710 ± 0.0071
	Loss Curvature [Garg et al.]	0.9827 ± 0.0019	0.9834 ± 0.0019	0.9849 ± 0.0014	0.9834 ± 0.0019
	CSL (Ours)	0.9867 ± 0.0007	0.9870 ± 0.0003	0.9869 ± 0.0007	0.9866 ± 0.0009
CIFAR-100	Thr. Learning Time (LT)	0.5144 ± 0.0017	0.5119 ± 0.0059	0.5069 ± 0.0050	0.5041 ± 0.0002
	In Conf. [Carlini et al.]	0.6706 ± 0.0052	0.6493 ± 0.0075	0.6324 ± 0.0051	0.6257 ± 0.0044
	CL [Northcutt et al.]	0.7233 ± 0.1707	0.7030 ± 0.1565	0.6833 ± 0.1427	0.6662 ± 0.1289
	SSFT [Maini et al.]	0.8495 ± 0.0002	0.8358 ± 0.0008	0.8203 ± 0.0016	0.8043 ± 0.0061
	Loss Curvature [Garg et al.]	0.9886 ± 0.0009	0.9887 ± 0.0013	0.9885 ± 0.0016	0.9888 ± 0.0004
	CSL (Ours)	0.9898 ± 0.0003	0.9897 ± 0.0003	0.9899 ± 0.0003	0.9899 ± 0.0002

Memorization & CSL vs Adversarial Distance (L2) — Fig 7: \( L_2 \) Adversarial distance vs Mem. score (Feldman & Zhang 2020) and CSL on CIFAR-100. Clearly, memorization and CSL increases for less robust samples (i.e. lower adv distance \( \|x - x_p\|_2 \)). Thus CSL captures similar properties as memorization