Neuro-JS is a lightweight JavaScript library for creating and training neural networks using backpropagation. It is designed for simplicity and flexibility, making it easy to experiment with different network configurations directly in the browser.
- Layered Neural Networks: Fully customizable multilayer perceptrons.
- Built-in Activation Functions: Includes sigmoid, ReLU, leaky ReLU, tanh, and linear functions.
- Customizable Training: Adjustable learning rate and error thresholds.
- Browser-Friendly: Pure JavaScript implementation, no external dependencies.
- Manual and Automatic Weight Initialization: Supports user-defined weights or auto-initialization.
These instructions will get you a copy of the project up and running on your local machine for development and testing purposes. See deployment for notes on how to deploy the project on a live system.
You need to download Neuro-JS project.
git clone https://github.com/s-litvin/neuro-js.git
Include Neuro-JS to your html-page
<body>
...
<script type="text/javascript" src="neuron.js"></script>
<script type="text/javascript" src="perceptron.js"></script>
</body>
Or use index.html (demo) from neuro-js project directory.
-
Create perceptron (with learning rate and error threshold):
perceptron = new Perceptron(0.98, 0.001);
-
Add neurons using array of neuron numbers for each layer:
createLayers(neuronesCountForEachLayer, linkAutomatically);
perceptron.createLayers([{'size': 2}, {'size': 2}, {'size': 1}]);
Supported activation functions:
Sigmoid(default): f(x) = 1 / (1 + e^(-x))ReLU: f(x) = max(0, x)LeakyReLU: f(x) = x > 0 ? x : 0.01 * xTanh: f(x) = (e^x - e^(-x)) / (e^x + e^(-x))Linear: f(x) = x
activation function ('sigmoid' is default):
perceptron.createLayers([{'size': 2}, {'size': 2, 'activation': Cell::RELU}, {'size': 1, 'activation': Cell::LINEAR}]);
Available activation types:
- Cell.RELU
- Cell.LEAKYRELU
- Cell.SIGMOID
- Cell.TANH
- Cell.LINEAR
In Neuro-JS, you can manually link neurons instead of relying on automatic linking. To do this, set the second parameter in the createLayers method to false. By default, this parameter is true, enabling automatic linking.
Here is an example demonstrating manual neuron linking:
// Create layers with manual linking
perceptron.createLayers([
{ size: 3, activation: Cell.SIGMOID }, // Input layer
{ size: 2, activation: Cell.SIGMOID }, // Hidden layer
{ size: 1, activation: Cell.LINEAR } // Output layer
], false);
// Manually link neurons with specified weights
perceptron.link('x00', 'h10', 0.1); // Link input neuron x00 to hidden neuron h10 with weight 0.1
perceptron.link('x00', 'h11', -0.2); // Link input neuron x00 to hidden neuron h11 with weight -0.2
perceptron.link('x01', 'h10', 0.4);
perceptron.link('x01', 'h11', 0.3);
perceptron.link('x02', 'h10', -0.5);
perceptron.link('x02', 'h11', 0.6);
perceptron.link('h10', 'y20', 0.7); // Link hidden neuron h10 to output neuron y20 with weight 0.7
perceptron.link('h11', 'y20', -0.1); // Link hidden neuron h11 to output neuron y20 with weight -0.1
// Set input and output vectors
perceptron.setInputVector([1.0, 0.5, 0.2]); // Set input values
perceptron.setOutputVector([0.8]); // Set target output
// Perform forward pass and backpropagation
perceptron.forwardPass(); // Calculate outputs based on inputs
perceptron.backPropagation(); // Adjust weights to reduce error-
Input Neurons (x): Prefixed with
x, followed by the layer index (always 0) and the neuron index.
Example:x02means input neuron, layer 0, neuron 3. -
Hidden Neurons (h): Prefixed with
h, followed by the layer index and the neuron index.
Example:h10means hidden neuron, layer 1, neuron 1. -
Output Neurons (y): Prefixed with
y, followed by the layer index and the neuron index.
Example:y20means output neuron, layer 2, neuron 1.
To link neurons manually, use the link method with the following parameters:
- Source neuron ID: The ID of the neuron from which the link starts (e.g.,
x00). - Target neuron ID: The ID of the neuron to which the link connects (e.g.,
h10). - Weight: The weight of the connection.
This approach gives you fine-grained control over the architecture and weight initialization of the neural network.
- Set input and target output vectors.
perceptron.setInputVector([0.61, 0.12]);
perceptron.setOutputVector([0.91]);
Now you have simple perceptron:
--(input1)----(hidden1)
\/ \_(output1)__
/\ /
--(input2)----(hidden2)
Dropout is a regularization technique used to prevent overfitting in neural networks. During training, dropout randomly disables a fraction of neurons in a layer for each forward pass. This helps the network become more robust by forcing it to learn redundant representations.
- Dropout rate is the fraction of neurons to disable during training. For example, a rate of
0.5means 50% of neurons will be randomly deactivated in each forward pass. - In Neuro-JS, you can enable dropout by specifying a rate when configuring the perceptron:
perceptron.setDropoutRate(0.5); // Set dropout rate to 50%
- The dropout mechanism is applied only during the training phase, so don't forget to turn it off after training completed. During inference (when the network is making predictions), all neurons are active, and their outputs are scaled appropriately.
- Prevents overfitting by ensuring that no single neuron becomes overly dominant in the network.
- Encourages the network to generalize better to unseen data.
- Useful in tasks where the dataset is small or has significant variability.
In the demo provided with Neuro-JS, you can adjust the dropout rate in real-time during the training process. Use the buttons DR + and DR - in the interface to increase or decrease the dropout rate dynamically. This allows for quick experimentation to find the optimal configuration for your network.
For more examples, visit the test page. This page provides various test cases demonstrating the capabilities of Neuro-JS.
The examples shown on the test page are executed using the tests.js file included in the repository. You can explore these examples to understand different ways to configure and train neural networks using this library.
Check demo on GitHub page: https://s-litvin.github.io/neuro-js/
