A simple neural network for digit classification built entirely from scratch in Python — no ML frameworks, no libraries like PyTorch or TensorFlow.
All matrix operations are implemented via a custom MatrixLab engine.
This project is designed as an educational example of how neural networks work under the hood.
- Image preprocessing (PNG → matrix)
- Forward pass (sigmoid activation)
- Backpropagation
- Gradient descent optimizer
- Manual save/load of weights
- Training visualization (matplotlib)
- No external ML libraries — all math coded manually
| Layer | Size |
|---|---|
| Input | 784 (28x28 grayscale image) |
| Hidden Layer | 100 neurons |
| Output | 10 neurons (digit classification: 0-9) |
| Activation | Sigmoid |
1️⃣ starter.py → generates random weights and saves them to /model-1/
2️⃣ train_script.py → runs training loop for several epochs and plots error graph
3️⃣ agent.py → contains full implementation of forward pass, backpropagation and gradient descent