# CNN BatchNorm Comparison — MNIST

Comparing a baseline CNN against the same architecture with Batch Normalization added, on MNIST digit classification.

## Architecture

```
Conv2D(32, 3x3, stride 2) → Conv2D(64, 3x3, stride 2) → Flatten → Dense(10, softmax)
```

The BatchNorm version inserts `BatchNormalization` + separate `Activation('relu')` after each conv layer, instead of the activation being baked into the conv layer directly.

## Clean ablation

Both models use the **identical optimizer and learning rate** (SGD, lr=0.001) — Batch Normalization is the only variable that changed. This makes the result directly attributable to BatchNorm, unlike the [MLP regularization comparison](../mnist-mlp-regularization/README.md) in this repo where two variables changed at once.

## Results

| Model | Test loss (categorical crossentropy) |
|---|---|
| Baseline CNN | 0.289 |
| CNN + BatchNorm | 0.094 |

BatchNorm alone (same optimizer, same learning rate) cut test loss by roughly two-thirds — consistent with its known effect of stabilizing and accelerating convergence, especially under plain SGD at a conservative learning rate.

## Tech stack

TensorFlow / Keras, NumPy, matplotlib

## Notebooks

- [`conv-pred.ipynb`](conv-pred.ipynb) — baseline
- [`conv-pred-BN.ipynb`](conv-pred-BN.ipynb) — with BatchNorm