Backpropagator

Backpropagator project is a project made by me and my friend for understanding how backpropagation works by building a computational graph and automating the chain rule from the ground up.

📖 About

In this project, my friend and I implemented Backpropagation using only Python and the math library. While modern frameworks like PyTorch handle massive tensors on GPUs, Backpropagator works at the scalar level, making it the perfect educational tool to understand exactly how gradients flow through a neural network.

✨ Key Features

• Scalar-Level Autograd: A custom Value class that stores data and its derivative (gradient).
• Automated Chain Rule: Uses a topological sort to ensure gradients are propagated in the correct order.
• Full Neural Stack: Includes building blocks for Neuron, Layer, and MLP (Multi-Layer Perceptron) architectures.
• Mathematical Operations: Supports +, -, *, /, **, and activation functions like tanh and exp.
• PyTorch Verified: Includes benchmarks to verify our gradient calculations against industry-standard results.

Class Diagram