Initial commit: MLP regularization, CNN BatchNorm comparison, autoencoder denoising

docs/Autoencoder-Denoising.md

@@ -0,0 +1,19 @@
+# Autoencoder Image Denoising — MNIST
+
+## Setup
+
+A minimal fully-connected autoencoder: `Input(784) → Dense(32, relu) → Dense(784, sigmoid)`. Gaussian noise (factor 0.5) is added to MNIST digits, and the network is trained to reconstruct the *original clean* image from the noisy input — binary crossentropy loss between reconstruction and ground truth.
+
+## Why a 32-unit bottleneck matters
+
+Compressing a 784-pixel image down to 32 values and then reconstructing it forces the network to learn which features actually matter for representing a digit, rather than memorizing pixels. If the reconstruction still looks like a recognizable digit after that bottleneck, it's evidence the encoder learned a genuinely useful compressed representation.
+
+## Result
+
+![Original, noisy, and denoised digits](images/autoencoder-denoising-result.png)
+
+Despite the aggressive compression (784 → 32 → 784), the reconstructions are clearly recognizable digits — the noise is substantially removed and the core digit shape is preserved.
+
+## Notebook
+
+[`encoder-noisy-image.ipynb`](../autoencoder-denoising/encoder-noisy-image.ipynb)