CNN from Scratch — MNIST & Fashion-MNIST

A convolutional neural network built from scratch using only NumPy, trained and evaluated on both MNIST and Fashion-MNIST datasets. No PyTorch, no TensorFlow, just raw math.

---

What's in this project

This notebook implements every layer of a CNN by hand, including the forward pass and backpropagation:

• Conv3x3 - Convolutional layer with 3×3 filters
• MaxPool2 - 2×2 max pooling layer
• Softmax - Fully connected output layer with softmax activation
• Cross-entropy loss - Used as the training objective
• SGD - Stochastic gradient descent for weight updates

---

Architecture

Input (28×28 grayscale)
    ↓
Conv3x3 (8 filters)   →  26×26×8
    ↓
MaxPool2              →  13×13×8
    ↓
Softmax (FC layer)    →  10 classes

---

Results

Trained for 3 epochs on 1,000 samples, tested on 1,000 samples.

Dataset	Test Loss	Test Accuracy
MNIST	0.4878	83.5%
Fashion-MNIST	0.6655	75.7%

Fashion-MNIST is a harder benchmark (clothing categories vs. handwritten digits), so the lower accuracy is expected.

---

Tech Stack

Tool	Purpose
NumPy	All layer math and backprop
TensorFlow / Keras	Dataset loading only (mnist, fashion_mnist)
Matplotlib	Loss & accuracy plots

---

Key Concepts Demonstrated

• Manual implementation of convolution and max pooling
• Backpropagation through each layer from first principles
• Cross-entropy loss and softmax gradient derivation
• Comparing CNN performance across datasets of different difficulty