MLP From Scratch (Python)
A 2–4–1 Python MLP trained on XOR without a deep-learning framework; the repository uses NumPy, MpMath, and Matplotlib.
Purpose
This project builds a multilayer perceptron (MLP) in Python to learn XOR without a deep-learning framework. Its forward pass and backpropagation show how inputs become predictions and how errors update weights.
Model and training
The network has two inputs, four hidden neurons, and one output, using sigmoid activations. Normally initialized weights train on 80 XOR samples for 5,000 epochs. The repository uses NumPy, MpMath, and Matplotlib: “from scratch” means no deep-learning framework, not no libraries.
Result
The README reports a 0.003 learning rate, loss of 0.0002100859, and 97.87174999% accuracy. Plots show loss and accuracy during training. These describe one small experiment; the goal is to follow backpropagation in code, not claim general performance.