2.1 KiB
MNIST Classification
A full tour of classification problem types on the MNIST handwritten digit dataset — binary, multiclass, multilabel, and multioutput — with an emphasis on evaluating models honestly rather than trusting raw accuracy.
Why not just report accuracy?
About 90% of MNIST digits are not an 8. A classifier that always guesses "not 8" scores ~90% accuracy while learning nothing. This project compares models with ROC AUC and F1 instead, and validates every "is it learning something real" claim against a DummyClassifier baseline.
Linear vs ensemble: SGD vs Random Forest
Same binary task ("is it an 8?"), two very different models:
| Model | F1 | ROC AUC |
|---|---|---|
| SGD (linear) | 0.58 | 0.915 |
| Random Forest (ensemble) | 0.90 | 0.997 |
The gap illustrates a core limitation of linear models: SGD can only draw one straight decision boundary through 784-dimensional pixel space, while Random Forest's ensemble of trees carves non-linear regions that fit the actual shape of the data.
Where the multiclass model actually gets confused
Extending to all 10 digits, the normalized confusion matrix shows exactly which digits get mistaken for which — more useful than a single accuracy number:
Multioutput: predicting an entire image
Pushed the idea further — instead of predicting a label, predicting all 784 pixel values at once, framed as denoising a noisy digit:
 |
 |
 |
| Original | Noisy input | Model's denoised output |
Notebooks
classification.ipynb— original workclassification_notebook_polished.ipynb— AI-assisted documentation pass