🧠 MNIST Neural Network (from Scratch)

A fully-connected neural network built entirely from scratch using NumPy — no TensorFlow, PyTorch, or Keras.
This project trains a simple feedforward model to classify handwritten digits from the MNIST dataset.

🚀 Overview

This project demonstrates the core mechanics of a neural network implemented manually:

• Matrix-based forward and backward propagation
• ReLU activation and Cross-Entropy Loss
• Mini-batch training with a custom DataLoader
• Weight initialization, gradient updates, and learning rate tuning
• Model saving/loading and prediction visualization with Matplotlib
  It was developed as a learning exercise to understand how deep learning works under the hood.

⚙️ Requirements

To install dependencies:

pip install -r requirements.txt

📦 Dataset

This project uses the MNIST dataset from Kaggle:
👉 Kaggle MNIST Dataset
Download the .idx files from there and place them in the data/ folder.

Place the following files inside the data/ folder:

• train-images.idx3-ubyte
• train-labels.idx1-ubyte
• t10k-images.idx3-ubyte
• t10k-labels.idx1-ubyte

🧠 Model Architecture

The default network used in main.py:

Layer	Type	Input → Output
1	Dense	784 → 256
2	ReLU	-
3	Dense	256 → 128
4	ReLU	-
5	Dense	128 → 10

Each neuron learns via stochastic gradient descent using cross-entropy loss.

🏋️‍♂️ Training

To train a new model:

python main.py

Then choose:

Train or Not (y/n): y

• Trains the model for 12 epochs (by default)
• Saves weights as mnist_model_1.npz
• Prints per-epoch losses and accuracies

Example output:

Final Train Accuracy: 91.87%
Final Test Accuracy : 90.42%

🔍 Testing / Evaluation

To test a pre-trained model:

python main.py

Then choose:

Train or Not (y/n): n

This loads your saved weights, evaluates accuracy on the MNIST test set, and displays 25 random digit predictions (5×5 grid) with predicted and true labels.

📊 Results

With basic tuning (lr=0.005, batch_size=128, epochs=12):

• Training Accuracy: ~92%
• Test Accuracy: ~90% Each run may vary slightly due to random initialization.

🙌 Acknowledgements

• MNIST dataset: Yann LeCun (http://yann.lecun.com/exdb/mnist/) et al.
• Data loader: adapted from Hojjat Khodabakhsh’s Kaggle notebook (https://www.kaggle.com/code/hojjatk/read-mnist-dataset)
• Created for personal learning and educational purposes.