LiteGrad 🧠

A lightweight, zero-dependency Neural Network and Matrix math library built entirely from scratch in pure TypeScript.

I built this library to bridge the gap between the mathematics of electrical engineering (like matrix-based signal processing) and software-based machine learning. Instead of relying on massive black-box libraries like TensorFlow or PyTorch, LiteGrad exposes the raw matrix calculus and backpropagation algorithms that power multi-layer networks.

Features

• Zero Dependencies: Pure TypeScript implementation from the ground up.
• Custom Math Engine: Includes a bespoke Matrix class for linear algebra operations (dot products, transpositions, Hadamard products).
• Multi-Layer Perceptron (MLP): Configurable input, hidden, and output layers.
• Backpropagation: Implements gradient descent with configurable learning rates.
• Modern Activations: Supports Sigmoid, ReLU, and Tanh activation functions.
• Model Persistence: Serialize and deserialize trained networks to JSON for easy storage and loading.

Installation

npm install litegrad


import { NeuralNetwork } from 'litegrad';

// 1. Initialize: 2 Inputs, 4 Hidden Nodes, 1 Output Node
const nn = new NeuralNetwork(2, 4, 1);
nn.learningRate = 0.1;

// 2. The Training Data (XOR Logic Gate)
const trainingData = [
  { inputs: [0, 0], targets: [0] },
  { inputs: [0, 1], targets: [1] },
  { inputs: [1, 0], targets: [1] },
  { inputs: [1, 1], targets: [0] },
];

// 3. Train the Network
console.log("Training started...");
for (let i = 0; i < 50000; i++) {
  const data = trainingData[Math.floor(Math.random() * trainingData.length)];
  
  // The train method returns the Mean Squared Error (MSE) for loss tracking
  const loss = nn.train(data.inputs, data.targets);
}
console.log("Training complete!");

// 4. Make Predictions
console.log("Guess for [0, 0]:", nn.predict([0, 0])[0]); // Expect ~0
console.log("Guess for [1, 0]:", nn.predict([1, 0])[0]); // Expect ~1