Transformer–CNN Comparative Study
Feb 2025 – Jun 2025The Lab
CNNs vs. Vision Transformers vs. hybrids on a self-collected dataset, with explainability.
- Built and evaluated CNN (EfficientNet), Vision Transformer, and hybrid ConvNeXt models for image classification on a self-collected campus dataset.
- Used transfer learning, extensive augmentation, and ensembling to push performance in a data-starved regime.
- Applied SHAP and Grad-CAM explainability to assess model behaviour and robustness, not just accuracy.
- Recorded a full video walkthrough of the methods and results.
- PyTorch
- Vision Transformer
- Explainable AI
- SHAP
- Grad-CAM