The JavaScript neural network that recommends its own replacements

What it does

brain is a JavaScript library for feed-forward neural networks. You feed it labeled training patterns—arrays or hashes of numbers scaled from 0 to 1—and it trains in bulk until the error falls below a threshold or a maximum iteration count is reached. Once trained, it predicts outputs for new inputs in Node or the browser.

The interesting bit

The library’s most useful quirk is toFunction(), which compiles a trained network into a plain JavaScript function string you can copy and paste into any project. That turns a model into a zero-dependency artifact, which was a neat deployment trick for lightweight browser code. For Node, it also exposes a WriteStream so you can pipe training data incrementally rather than loading the entire dataset at once.

Key highlights

• Feed-forward architecture with configurable hidden layers
• Accepts training data as arrays or object hashes of normalized 0–1 values
• toFunction() exports a trained model as a standalone JS function string, removing the runtime dependency
• Node WriteStream support for streaming training data via pipe()
• 7,990 GitHub stars, though the author has ended active development

Caveats

• The author explicitly recommends against using it today; active development has ceased in favor of libraries like brain.js and convnetjs
• Standard training requires the entire dataset in a single bulk call unless you use the streaming API
• Training is computationally expensive and expected to run offline or in a Web Worker

Verdict

Browse the code if you are curious about early JavaScript machine-learning patterns or maintaining a legacy system. If you are starting a new project, follow the author’s own advice and look elsewhere.