Using a pre-trained ResNet-50 or Vision Transformer (ViT) to extract the embedding vector for 148_1000.jpg .
Applying t-SNE or UMAP to see where this image sits relative to its assigned class.
(e.g., ImageNet, a local project, or a specific website?) 148_1000.jpg
Generating Grad-CAM visualizations to identify which pixels the model focuses on when classifying this specific image. 3. Results & Discussion
The rise of deep learning relies on massive datasets where individual image quality and annotation accuracy are often assumed rather than verified. Using a pre-trained ResNet-50 or Vision Transformer (ViT)
Edge cases or "noisy" samples (like 148_1000.jpg ) can disproportionately affect model convergence or bias.
(e.g., An animal, a vehicle, a medical scan?) a local project
Recommendations for automated "cleaning" of datasets based on high-loss samples.