A fully interactive multilayer neural network built in pure Python and NumPy, capable of training via backpropagation, saving/loading trained models, and classifying both binary and multi-class datasets.
This project implements a Multilayer Perceptron (MLP) from scratch, without the use of any specialized machine learning frameworks. The network supports configurable architecture, iterative training with backpropagation, forward-pass inference, and persistent model storage — all through a console-based interface.
This project builds upon a single-layer perceptron implementation, extending it into a full deep learning pipeline with automatic differentiation via backpropagation.
- Python 3.x — core language
- NumPy — matrix operations, weight initialization, normalization
- Matplotlib — scatter plots, error curves, confusion matrix
- No PyTorch, TensorFlow, or Keras — fully manual implementation
- 🏗️ Fully configurable network architecture (inputs, hidden layers, neurons per layer, outputs)
- 🔄 Backpropagation training with configurable learning rate and epochs
- ➡️ Forward-pass inference on unseen test data
- 💾 Save and load trained models to/from CSV files
- 📊 Automatic generation of 3 visualizations after training and testing:
- Scatter plot of real data (2D or 3D projection)
- Correct vs incorrect predictions per data point
- Training: error evolution per epoch · Testing: confusion matrix
- 🔢 Supports binary and multi-class classification (One-Hot Encoding)
- 🔁 Continuous training loop — keep training the same model across multiple sessions
multilayer-perceptron/
│
├── main.py # Entry point and main menu logic
├── neural_network.py # Network core: initialization, forward pass, backpropagation
├── activation_functions.py # Activation functions (Sigmoid + derivative)
├── data.py # CSV loading, normalization, One-Hot encoding, save/load model
├── plots.py # Visualizations: scatter plots, error curve, confusion matrix
│
├── 2d_2_class_train.csv # Training set: 2D input, 2-class output
├── 2d_2_class_test.csv # Test set: 2D input, 2-class output
├── trained_model_2d_2.csv # Trained model: 2D / 2-class (ready to load)
├── 3d_4_class_train.csv # Training set: 3D input, 4-class output
├── 3d_4_class_test.csv # Test set: 3D input, 4-class output
└── trained_model_3d_4.csv # Trained model: 3D / 4-class (ready to load)
| Plot | Description |
|---|---|
| 1. Real Data | Scatter plot colored by true class label |
| 2. Predictions | Green = correct · Red = incorrect |
| 3. Error Curve | Mean squared error per epoch |
| Plot | Description |
|---|---|
| 1. Real Data | Scatter plot colored by true class label |
| 2. Predictions | Green = correct · Red = incorrect |
| 3. Confusion Matrix | Predicted vs actual class using matshow |
For datasets with 3+ input dimensions, the first three are rendered using a 3D projection.
🚀 Getting Started
git clone https://github.com/iarondon3/multilayer-perceptron.git
cd multilayer-perceptronpip install numpy matplotlibpython main.py1. Choose: create a new network or load a saved model
2. If new → configure: inputs, outputs, hidden layers, neurons per layer
3. Main menu:
├── 1. Train → provide training CSV + number of epochs
├── 2. Test → provide test CSV → see predictions + confusion matrix
├── 3. Save → export trained model to CSV
└── 4. Exit
4. After training or testing, visualizations are displayed automatically
5. Return to main menu to continue training or testing
📁 File Formats
One vector per row — last column is the class label:
x1,x2,class
0.52,0.91,1
0.13,0.47,-1
...Trained models are saved with hyperparameters in the first 5 rows, followed by one row per neuron:
u, 2 ← input size
v, 1 ← output size
L, 2 ← hidden layers
b, 4 ← neurons per layer
e, 5000 ← epochs trained
h,1,0,w1,...,bias
o,2,0,w1,...,bias
🧮 Mathematical Foundation
Forward Pass (per layer):
Z = X · W + b
A = σ(Z)
Backpropagation:
δ_output = (target - output) · σ'(output)
δ_hidden = (δ_next · Wᵀ) · σ'(activation)
W = W + α · Aᵀ · δ
Activation Function — Sigmoid:
σ(x) = 1 / (1 + e^(-x))
σ'(x) = σ(x) · (1 - σ(x))
Multi-class support via One-Hot Encoding — automatically applied when more than 2 classes are detected in the dataset.
Isabella Rondón | Business Economist & Data Analyst
