Neural Network From Scratch 🧠

This project implements a simple feedforward neural network from scratch in Python, without using high-level machine learning libraries like TensorFlow, PyTorch, or NumPy. It is designed for educational purposes and demonstrates the core concepts of neural networks, including forward propagation, backpropagation, and training on real-world data.

Features ⚙️

• Custom Matrix and Vector classes for all linear algebra operations (see my_tools/VectorMatrixClass.py)
• Manual implementation of:
  • Forward propagation
  • Backpropagation
  • Weight and bias updates
  • Cost calculation (binary cross-entropy)
  • Early stopping (optional)
• Data loading from CSV (Breast Cancer Wisconsin dataset)
• Configurable network architecture (number of layers, neurons per layer)
• Training and validation split
• Visualization of training cost over epochs
• Validation accuracy reporting

File Structure 📂

Neural_Network_FromScratch/
├── archive/
│   └── data.csv                # Breast Cancer dataset (CSV)
├── my_tools/
│   ├── VectorMatrixClass.py    # Custom Matrix and Vector classes
│   └── my_random.py            # Random number utilities
├── neuron.py                   # Main neural network implementation

How to Run ▶️

1. Install requirements
   Only standard Python libraries are used (no external dependencies).

2. Prepare the data
   Ensure archive/data.csv is present (already included).

3. Run the main script

   python3 neuron.py

4. What you will see

   • Training cost plotted over epochs
   • Final validation accuracy printed

Main Components 🧩

• neuron.py
  Main script: loads data, initializes the network, trains, validates, and plots results.

• my_tools/VectorMatrixClass.py
  Custom classes for matrix and vector operations (addition, multiplication, transpose, etc.).

• my_tools/my_random.py
  Utilities for generating random numbers for weight initialization.

• archive/data.csv
  Breast Cancer Wisconsin dataset (features and labels).

Customization 🛠️

• Network architecture:
  Change layer_nb and the starting number of neurons in init_neuron_layers().

• Learning rate and epochs:
  Adjust alpha and epochs in main().

• Early stopping:
  The train() function supports early stopping via patience and min_delta parameters.

Example Output 📋

[30, 64, 32, 16, 1]
455 samples in training set
epoch 0: cost = 0.693147
...
Final validation accuracy: 0.95

Notes 📝

• No external ML libraries are used; all math is implemented manually.
• For best results, consider normalizing your input features.
• This project is for learning and demonstration; for production use, prefer established libraries.
• With fast parameters, the accuracy is ~65%
• I’ve been able to increase the accuracy to 85–90% with other parameters, but the whole process becomes very slow.
• Feel free to tweak the parameters as you like to try having a better accuracy !

---

Author:
Dorian Boiré

2025