Neural Network from Scratch 🧠

A fully functional neural network implementation built entirely from NumPy, featuring:

• ✅ Forward and backward propagation
• ✅ Adam optimizer with bias correction
• ✅ Multiple activation functions (ReLU, tanh, sigmoid, softmax)
• ✅ Proper weight initialization (Xavier/He)
• ✅ Numerical stability (clipping, softmax stabilization)
• ✅ Batch training with shuffling

Results

XOR Problem

• 100% accuracy - solved the classic non-linear problem

MNIST Handwritten Digits

• 96.07% test accuracy with just 10k training samples
• Architecture: 784 → 128 → 64 → 10

Installation

pip install -r requirements.txt

Usage

XOR Problem Example

python examples/xor_example.py

MNIST Classification Example

python examples/mnist_example.py