A minimal neural network built entirely from scratch using NumPy — designed to learn the classic XOR logic function.
Note
This project is experimental and intended for learning purposes only.
To demystify how neural networks work under the hood by building one from scratch —
no libraries like TensorFlow or PyTorch — just pure matrix math using NumPy.
This project solves the XOR problem, which is not linearly separable (i.e., it can't be solved by a straight line).
We use 1 hidden layer with 2 neurons, and sigmoid activations.
- Understand forward and backward propagation
- Learn how weights and biases are updated
- See how activation functions help neural nets learn
- Build confidence to create deeper nets from scratch
| Input (X1, X2) | Output |
|---|---|
| 0, 0 | 0 |
| 0, 1 | 1 |
| 1, 0 | 1 |
| 1, 1 | 0 |
| Section | What It Does |
|---|---|
X, Y |
Input/Output values for XOR |
sigmoid() |
Activation function |
derivative() |
Used in backprop to adjust weights |
w1, w2, b1, b2 |
Random weights and biases for layers |
| Forward Pass | Calculates output |
| Backward Pass | Calculates how wrong and how to fix it |
| Weight Update | Adjusts weights and biases to improve accuracy |
| Output | Final predictions after training |
- Python 3.10+
- NumPy
...
git clone (https://github.com/Elitsuv/NeuXor.git)
cd .
python network.py