MNIST

A notebook-driven analysis of MNIST treated as a statistical dataset, not just a classification benchmark. Six ordered notebooks go from raw acquisition to final comparison: validation and loading, preprocessing, exploratory analysis of pixel distributions and per-digit behavior, a multinomial logistic regression baseline evaluated with accuracy, balanced accuracy, macro and weighted precision-recall-F1, log loss, and top-3 accuracy, then a repeated-sampling study (30 runs at 10,000 samples) that quantifies how much accuracy varies between random draws, and finally PCA-based dimensionality reduction compared against the baseline features. Trained models, metrics, and generated plots are exported so every step is reproducible.