CNN MNIST Classifier

Demo samples are pre-packaged by class 0..9. The base load pulls 10 images per class (100 total).

Use Add +10 per Class to incrementally scale to 20, 30, 40, 50 per class (max 500 total).

This setup is designed for fast educational experiments while preserving class balance.

The model outputs a 10-way softmax distribution.

$$\hat{y}_c = \frac{e^{z_c}}{\sum_{k=0}^{9} e^{z_k}}, \quad L = -\frac{1}{N}\sum_i\sum_{c=0}^{9} y_{ic}\log(\hat{y}_{ic})$$

Use learning rate, batch size, optimizer, and architecture options to trade off speed and stability.