cazala/synaptic

What it solves

Synaptic provides a flexible way to build and train neural networks using JavaScript, allowing them to run in both Node.js environments and directly in the web browser. It removes the restriction of fixed architectures, enabling developers to create custom first-order and second-order neural network structures.

How it works

The library uses a generalized training algorithm based on Derek D. Monner's research for second-order recurrent neural networks. It provides a set of primitive building blocks—Neurons, Layers, and Networks—that can be connected to form various architectures. It also includes a Trainer class to handle the learning process and an Architect class that provides pre-defined common architectures.

Who it’s for

It is designed for JavaScript developers who want to implement neural networks in their applications without needing a heavy backend or specialized AI hardware, as well as those experimenting with different neural network topologies.

Highlights

• Architecture-free: Supports the creation of virtually any neural network type, including custom recurrent structures.
• Cross-platform: Works seamlessly in both Node.js and the browser.
• Built-in Architectures: Includes ready-to-use implementations of Multilayer Perceptrons, LSTMs, Liquid State Machines, and Hopfield networks.
• Integrated Testing: Comes with a trainer capable of running standard tests like XOR and Distracted Sequence Recall to compare network performance.