Building a Multilayer Perceptron from Scratch: Classifying MAGIC Gamma Ray Data with Pure Python

This project was developed by mhenry27, which builds a Multilayer Perceptron (MLP) entirely from scratch using Python for classifying the MAGIC Gamma Telescope dataset. It includes complete implementations of forward propagation, backpropagation, gradient descent, and hyperparameter tuning, aiming to help learners deeply understand the underlying mechanisms of neural networks and having high educational value.

• Original Author/Maintainer: mhenry27
• Source Platform: GitHub
• Original Project Name: Neural-Network-Project
• Original Link: https://github.com/mhenry27/Neural-Network-Project
• Release Date: June 12, 2026

The MAGIC (Major Atmospheric Gamma Imaging Cherenkov) telescopes are two atmospheric Cherenkov telescopes located on La Palma Island in the Canary Islands, used to detect high-energy cosmic gamma rays. Distinguishing gamma ray signals from background cosmic ray noise is a key task in astrophysics.

The MAGIC dataset contains features of gamma ray signals and background cosmic ray noise, such as ellipticity, size, asymmetry, etc. It is a classic machine learning benchmark dataset with characteristics of real-world scientific data (e.g., need for feature scaling, class imbalance, etc.).

The biggest highlight of the project is that it does not rely on any deep learning framework, and implements core components manually:

1. Forward Propagation: Implements matrix operation processes, including linear transformation (weight matrix + bias vector) and application of nonlinear activation functions, where information flows from the input layer to the output layer.
2. Backpropagation: Calculates gradients based on the chain rule, propagating error signals layer by layer from the output layer to the input layer, which is the key to network learning.
3. Gradient Descent Optimization: Implements the basic gradient descent algorithm, minimizing the loss function by iteratively updating weights and biases, laying the foundation for understanding complex optimizers.

The project conducts systematic research on key hyperparameters:

• Hidden Layer Size: Experiments on the impact of different numbers of neurons; too small is prone to underfitting (unable to capture complex patterns), while too large may lead to overfitting or reduced training efficiency.
• Learning Rate: Analyzes the impact of different settings on convergence speed and final performance; a too large learning rate may cause oscillations and non-convergence, while a too small one leads to slow convergence.

This project has significant value for learners:

• Principle Understanding: By implementing components manually, learners can deeply understand the working principles of neural networks, rather than just staying at the API calling level.
• Debugging Ability: Handling numerical issues (such as gradient vanishing/explosion) cultivates practical debugging skills.
• Performance Optimization: Understanding computation graphs and memory usage lays the foundation for efficient implementation.
• Research Foundation: After mastering the underlying mechanisms, it becomes easier to understand and implement the latest research technologies.

The project's code structure is clear and has detailed comments, making it suitable for teaching:

• Efficient implementation of matrix operations
• Support for multiple activation functions (e.g., Sigmoid, ReLU, etc.)
• Loss function calculation and gradient derivation
• Complete implementation of training loop
• Calculation of model evaluation metrics

In an era where deep learning frameworks are highly abstract, this project reminds us that understanding basic principles is the necessary path to mastering advanced technologies. Whether for interview preparation, academic research, or knowledge pursuit, building a neural network from scratch is a worthwhile practice.

The project demonstrates how machine learning can assist astrophysics research, helping scientists identify valuable gamma ray signals from massive amounts of data.