Visual and Quantitative Analysis of Deep Learning Robustness to Shear in Mechanical MNIST

ABSTRACT

This study quantifies deep neural network robustness to physically induced distortions using MNIST and Mechanical MNIST (Step 5), a displacement‐field variant. A custom two‐layer convolutional neural network (CNN) with 32 and 64 filters attains 99.18% accuracy on MNIST and 98.17% on Mechanical MNIST, outperforming ResNet‐18, employed here as a representative deep residual baseline, which drops from 98.95% on MNIST to 83.46% on Mechanical MNIST, suggesting that within this controlled benchmark setting, compact domain‐aligned architectures demonstrate greater robustness to mechanical perturbation than deeper residual networks. Grad‐CAM class activation maps reveal diffuse, spatially dispersed activation patterns under mechanical perturbation, particularly for curvilinear digits 3, 5, and 8. These findings highlight the fragility of deep residual architectures relative to compact domain‐aligned models under controlled mechanical perturbation, motivating future investigation of domain‐adaptive architectures in physically perturbed classification domains such as tactile sensing and structural analysis.