Neural Network from Scratch

How it works

This neural network was built from scratch using only NumPy - no TensorFlow, PyTorch, or other deep learning frameworks for inference.

• Architecture: 784 → 128 → 64 → 10 neurons
• Trained on 60,000 MNIST handwritten digits
• Achieves ~98% accuracy on test data
• Forward pass implemented with pure NumPy matrix operations

Technical Details

PyTorch was used only for training (GPU acceleration, Adam optimizer). The trained weights were exported and loaded into the from-scratch NumPy implementation.