Multi-sample $$\zeta $$ -mixup: richer, more realistic synthetic samples from a p-series interpolant

Journal of Big Data

Table 5 Classification error rate (ERR) improvements on CIFAR-10 and CIFAR-100 datasets with $\zeta $-mixup applied in conjunction with CutMix

Method	ResNet-18	ResNet-50	MobileNetV2	EfficientNet-B0
CIFAR-10
CutMix	$4.13 \pm 0.01$	$4.08 \pm 0.12$	$8.97 \pm 0.08$	${9.99 \pm 0.29}$
+ $\zeta $-mixup	$\varvec{3.84 \pm 0.08}$ + *7.02%*	$\varvec{3.61 \pm 0.06}$ + *11.52%*	$\varvec{8.18 \pm 0.09}$ + *8.81%*	$\varvec{9.15 \pm 0.08}$ + *8.41%*
CIFAR-100
CutMix	${19.97 \pm 0.07}$	$18.99 \pm 0.08$	$28.93 \pm 0.18$	$31.55 \pm 0.15$
+ $\zeta $-mixup	$\varvec{19.54 \pm 0.06}$ + *2.15%*	$\varvec{18.86 \pm 0.04}$ + *0.68%*	$\varvec{28.31 \pm 0.25}$ + *2.14%*	$\varvec{30.73 \pm 0.07}$ + *2.29%*

The lowest errors are formatted with bold. Percentage relative improvements over using only CutMix are shown in green. ERRs are reported as mean ± standard deviation over 3 runs. Lower values are better

Method	ResNet-18	ResNet-50	MobileNetV2	EfficientNet-B0
CIFAR-10
CutMix	\(4.13 \pm 0.01\)	\(4.08 \pm 0.12\)	\(8.97 \pm 0.08\)	\({9.99 \pm 0.29}\)
+ \(\zeta \)-mixup	\(\varvec{3.84 \pm 0.08}\) + *7.02%*	\(\varvec{3.61 \pm 0.06}\) + *11.52%*	\(\varvec{8.18 \pm 0.09}\) + *8.81%*	\(\varvec{9.15 \pm 0.08}\) + *8.41%*
CIFAR-100
CutMix	\({19.97 \pm 0.07}\)	\(18.99 \pm 0.08\)	\(28.93 \pm 0.18\)	\(31.55 \pm 0.15\)
+ \(\zeta \)-mixup	\(\varvec{19.54 \pm 0.06}\) + *2.15%*	\(\varvec{18.86 \pm 0.04}\) + *0.68%*	\(\varvec{28.31 \pm 0.25}\) + *2.14%*	\(\varvec{30.73 \pm 0.07}\) + *2.29%*