Visual and Quantitative Analysis of Deep Learning Robustness to Shear in Mechanical MNIST
DOI:
https://doi.org/10.56147/aaiet.1.5.95Keywords:
- Convolutional neural networks,
- Image classification,
- Shear perturbations,
- Mechanical MNIST,
- Feature extraction
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 3-layer Convoluted Neural Network, CNN (32–256 filters) attains 98.17% accuracy on Mechanical MNIST, surpassing ResNet-18, which falls from 98.95% (MNIST) to 83.46%. Grad-CAM exposes saliency degradation under mechanical stress, notably for curved digits (‘3’, ‘7’); t-SNE and noise sensitivity analyses reveal distorted embeddings and diminished class separability. These findings highlight the fragility of deep, generic architectures in mechanically perturbed domains, advocating compact, domain-adaptive models for resilient classification in applications like tactile sensing or structural analysis.
