Toy Neural Network JS Library

This is a basic JavaScript neural network library. It can be trained in the browser and the resulting network can be saved to JSON for later deployment. It utelized a scratch-written matrix math class mostly for learning purposes and is by no means effiecient or fast.

Usage

A network can be initialized by creating a new NeuralNetwork object with 3 parameters: number of input nodes, hidden nodes, and output nodes.

brain = new NeuralNetwork(784, 64, 4);

The network can be trained using the train() function and passing it an array of input values and the label (also passed as an array with one-hot). The input values should be normalized to the range 0-1.

inputs = [0.23512, 0.8982053, ..., 0.123452];
target = [0, 0, 1, 0];
brain.train(inputs, target);

This training should be done many times. I found it best to pass all of the training data, randomize the order of the training data, and then pass it again. Around 5 of these epochs seemed to yeild good results.

After training, a prediction can be made on new data using the predict() function.

let prediction = brain.predict(input);

This will return an array of values between 0 and 1 corresponding to the labels. The highest value corresponds to the highest likely class given that input.

Saving and Using a Trained Network

Once trained, the network can be exported to JSON using the serialize() function. I belive this will trigger a download prompt in most browsers to save the JSON file.

brain.serialize();

Later on when you would like to utilize this saved network, simply load the JSON file in your preload, and in your setup call the static function deserialize() passing the JSON.

let brain = NeuralNetwork.deserialize(nn_json_object);

Future of Project

The main thing I would like to do next with this library is encorporate cpu/gpu parallelization for the matrix math. This would increase training performance greatly. Currently training performance is prohibitively poor. Encorporating a small library like gpu.js would be interesting and a great way to boost performance.

I would like to learn more about encorporating convolutional layers into the network for using images as input.

Any questions or comments send them my way!