TITLE:
Less Is More: When Data Augmentation Hurts Small CNN Generalization
AUTHORS:
Rui Huang
KEYWORDS:
Data Augmentation, Small CNNs, Low-Data Learning, Generalization, CIFAR-10, FashionMNIST, Model Capacity, Underfitting
JOURNAL NAME:
Journal of Computer and Communications,
Vol.14 No.7,
July
7,
2026
ABSTRACT: Data augmentation is widely adopted as a default regularization strategy for training deep neural networks, yet its effectiveness in low-data regimes with capacity-constrained models remains poorly characterized despite its practical importance. We present a systematic empirical study of one cumulative augmentation hierarchy for small convolutional neural networks trained on limited subsets of CIFAR-10 and FashionMNIST. The study contains 35 experimental conditions, each evaluated with ten fixed random seeds, for 350 total model fits. These conditions cover seven progressively aggressive augmentation levels for CIFAR-10 SmallCNN at three training-set sizes, CIFAR-10 TinyCNN at
n=1000
, and FashionMNIST SmallCNN at
n=1000
. Augmentation beyond a single horizontal flip consistently and substantially degrades test accuracy in this hierarchy. On CIFAR-10 with 500 training samples, maximum augmentation reduces accuracy by 10.9 percentage points below the no-augmentation baseline, and on 5000 samples the degradation reaches 16.6 points. Statistical analysis confirms that the high-augmentation degradations are significant after Bonferroni correction (paired t-test,
p0.001
, Cohen’s
d>4.0
) and appear across architectures and datasets. Our analysis reveals that aggressive augmentation causes a transition from overfitting to underfitting of the augmented training distribution. These findings challenge the common assumption that stacking augmentation transforms necessarily improves generalization and provide actionable guidance for practitioners working with limited labeled data and constrained model budgets in edge AI, medical imaging, and domain-specific applications.