Continuous-Time Networks Library

A JavaScript library for simulating continuous-time dynamical networks, with support for browser and Node.js environments. This library implements CTRNNs (Continuous-Time Recurrent Neural Networks) with TensorFlow.js for fast simulation.

Features

• Fast simulation of continuous-time networks using TensorFlow.js
• Support for multiple computation backends (CPU, WebGL, WASM)
• Browser and Node.js compatibility
• Async generator-based API for streaming simulation results
• Built-in oscillator examples
• Comprehensive testing across environments

Status

This library is experimental and in active development. It's published to npm.

Installation

The package is available on npm:

# Install globally
npm install -g @mazeballs/ctnet

# Or install in a project
npm install @mazeballs/ctnet --save-dev

Alternatively, you can install directly from the repository:

git clone https://github.com/danbri/mazeballs.git
cd mazeballs/continuous-time-networks
npm install

Quick Start

Node.js

// Import CTNet
const { CTNet } = require('@mazeballs/ctnet');

// TensorFlow.js is needed 
const tf = require('@tensorflow/tfjs');

// Create a two-node oscillator
const net = CTNet({
  size: 2,
  init_weights: [
    [4.5, 1],
    [-1, 4.5]
  ]
});

// Configure the network (two-step initialization)
net.states = tf.tensor1d([0.1, 0.5]);
net.biases = tf.tensor1d([-2.75, -1.75]);
net.step_size = 0.01;
net.run_duration = 1000;

// Run simulation using async generator
(async () => {
  for await (const result of net.runSimulation()) {
    console.log(result.outputs_cpu);
    // Access result.states, result.outputs, result.yprime
  }
})();

Browser

<script src="https://cdn.jsdelivr.net/npm/@tensorflow/tfjs"></script>
<script src="dist/ctnet.js"></script>
<script>
  // Create a basic oscillator
  const net = CTNet({
    size: 2,
    init_weights: [
      [4.5, 1],
      [-1, 4.5]
    ]
  });
  
  // Configure the network
  net.states = tf.tensor1d([0.1, 0.5]);
  net.biases = tf.tensor1d([-2.75, -1.75]);
  net.step_size = 0.01;
  net.run_duration = 1000;
  
  // Run simulation asynchronously
  async function runSimulation() {
    for await (const result of net.runSimulation()) {
      console.log(result.outputs_cpu);
    }
  }
  
  runSimulation();
</script>

Two-Step Initialization Pattern

This library uses a deliberate two-step initialization process:

1. Create the network with initial configuration (weights, size)
2. Manually set other parameters after initialization

// Step 1: Create network with basic structure
const myNet = CTNet({
  size: 2,
  init_weights: [
    [4.5, 1],
    [-1, 4.5]
  ]
});

// Step 2: Configure runtime parameters
myNet.states = tf.tensor1d([0.1, 0.5]);
myNet.biases = tf.tensor1d([-2.75, -1.75]);
myNet.step_size = 0.01;
myNet.run_duration = 1000;

Examples

See the examples directory for more detailed usage examples:

• Node.js examples in examples/node
• Browser examples in examples/browser

Demos

Interactive demonstrations are available in the demos directory.

TensorFlow.js Backends

This library supports multiple TensorFlow.js computation backends:

• WebGL (default in browser environments)
• WASM (WebAssembly - good cross-platform performance)
• CPU (always available)
• WebGPU (experimental, requires browser support)

You can select a backend using:

// Import the library with integrated backend support
const { CTNet, backends } = require('@mazeballs/ctnet');

// Backends are automatically set up with default preference: wasm > webgl > cpu
// You can check what backend was selected:
console.log('Current backend:', backends.getCurrentBackend());

// List all available backends
const availableBackends = backends.getRegisteredBackends();
console.log('Available backends:', availableBackends);

// You can also customize the backend preference order:
await backends.setBestAvailableBackend(['cpu', 'wasm', 'webgl']);

// Or directly on a network instance:
const net = CTNet({ size: 2 });
await net.setBackendPreferences(['wasm', 'cpu']);
console.log('Network using backend:', net.getBackend());

To verify WASM is working correctly, run the built-in utility:

npm run test:wasm

License

Apache License 2.0