Convolutional neural networks that train in your browser

What it does

ConvNetJS is a JavaScript library for building and training neural networks entirely in the browser. It supports fully-connected and convolutional layers, plus classification, regression, and an experimental reinforcement-learning module. A suite of live demos lets you watch MNIST, CIFAR-10, and toy datasets train in real time.

The interesting bit

Authored by Andrej Karpathy, the library treats the browser as a first-class deep-learning runtime rather than just a visualization shell. You assemble models by pushing layer definitions into an array, which gives the API a declarative feel despite being built on plain JavaScript.

Key highlights

• Handles convolutions, pooling, fully-connected layers, ReLU, softmax, SVM, and L2 regression
• Includes an experimental Deep Q Learning module for browser-based reinforcement learning
• Live demos cover image classification, autoencoders, 2D decision boundaries, and optimizer comparisons
• Represents tensors as 3-D volumes (sx, sy, depth) with a helper to convert DOM images into model inputs
• Also targets Node.js, though the author suspects the npm package is currently broken

Caveats

• Explicitly unmaintained: the author has no time for updates or bug fixes
• The npm package may not work, so Node users should expect rough edges

Verdict

Grab it if you want a self-contained, no-backend sandbox for learning neural-net internals. Look elsewhere for production training pipelines or actively maintained libraries.