Synapses

A neural networks library for Elixir!

# add
{:synapses, "~> 7.4.1"}
# to mix.exs

Neural Network

Create a neural network

Import Synapses, call NeuralNetwork.init and provide the size of each layer.

alias Synapses.{ActivationFunction, NeuralNetwork, DataPreprocessor, Statistics}
layers = [4, 6, 5, 3]
neuralNetwork = NeuralNetwork.init(layers)

neuralNetwork has 4 layers. The first layer has 4 input nodes and the last layer has 3 output nodes. There are 2 hidden layers with 6 and 5 neurons respectively.

Get a prediction

inputValues = [1.0, 0.5625, 0.511111, 0.47619]
prediction =
  NeuralNetwork.prediction(neuralNetwork, inputValues)

prediction should be something like [ 0.8296, 0.6996, 0.4541 ].

Note that the lengths of inputValues and prediction equal to the sizes of input and output layers respectively.

Fit network

learningRate = 0.5
expectedOutput = [0.0, 1.0, 0.0]
fitNetwork =
  NeuralNetwork.fit(
    neuralNetwork,
    learningRate,
    inputValues,
    expectedOutput
  )

fitNetwork is a new neural network trained with a single observation.

To train a neural network, you should fit with multiple datapoints

Create a customized neural network

The activation function of the neurons created with NeuralNetwork.init, is a sigmoid one. If you want to customize the activation functions and the weight distribution, call NeuralNetwork.customizedInit.

activationF = fn (layerIndex) ->
  case layerIndex do
    0 -> ActivationFunction.sigmoid
    1 -> ActivationFunction.identity
    2 -> ActivationFunction.leakyReLU
    _ -> ActivationFunction.tanh
  end
end

weightInitF = fn (_layerIndex) ->
  1.0 - 2.0 * :rand.uniform()
end

customizedNetwork =
  NeuralNetwork.customizedInit(
    layers,
    activationF,
    weightInitF
  )

Visualization

Call NeuralNetwork.toSvg to take a brief look at its svg drawing.

The color of each neuron depends on its activation function while the transparency of the synapses depends on their weight.

svg = NeuralNetwork.toSvg(customizedNetwork)

Save and load a neural network

JSON instances are compatible across platforms! We can generate, train and save a neural network in Python and then load and make predictions in Javascript!

toJson

Call NeuralNetwork.toJson on a neural network and get a string representation of it. Use it as you like. Save json in the file system or insert into a database table.

json = NeuralNetwork.toJson(customizedNetwork)

ofJson

loadedNetwork = NeuralNetwork.ofJson(json)

As the name suggests, NeuralNetwork.ofJson turns a json string into a neural network.

Encoding and decoding

One hot encoding is a process that turns discrete attributes into a list of 0.0 and 1.0. Minmax normalization scales continuous attributes into values between 0.0 and 1.0. You can use DataPreprocessor for datapoint encoding and decoding.

The first parameter of DataPreprocessor.init is a list of tuples (attributeName, discreteOrNot).

setosaDatapoint = %{
  "petal_length" => "1.5",
  "petal_width" => "0.1",
  "sepal_length" => "4.9",
  "sepal_width" => "3.1",
  "species" => "setosa"
}

versicolorDatapoint = %{
  "petal_length" => "3.8",
  "petal_width" => "1.1",
  "sepal_length" => "5.5",
  "sepal_width" => "2.4",
  "species" => "versicolor"
}

virginicaDatapoint = %{
  "petal_length" => "6.0",
  "petal_width" => "2.2",
  "sepal_length" => "5.0",
  "sepal_width" => "1.5",
  "species" => "virginica"
}

dataset = [setosaDatapoint, versicolorDatapoint, virginicaDatapoint]

dataPreprocessor =
  DataPreprocessor.init(
    [
      {"petal_length", false},
      {"petal_width", false},
      {"sepal_length", false},
      {"sepal_width", false},
      {"species", true}
    ],
    dataset
  )
  
encodedDatapoints = Enum.map(
  dataset,
  fn x ->
    DataPreprocessor.encodedDatapoint(dataPreprocessor, x)
  end
)

encodedDatapoints equals to:

[ [ 0.0     , 0.0     , 0.0     , 1.0     , 0.0, 0.0, 1.0 ],
  [ 0.511111, 0.476190, 1.0     , 0.562500, 0.0, 1.0, 0.0 ],
  [ 1.0     , 1.0     , 0.166667, 0.0     , 1.0, 0.0, 0.0 ] ]

Save and load the preprocessor by calling DataPreprocessor.toJson and DataPreprocessor.ofJson.

Evaluation

To evaluate a neural network, you can call Statistics.rootMeanSquareError and provide the expected and predicted values.

expectedWithOutputValues =
  [
    {[0.0, 0.0, 1.0], [0.0, 0.0, 1.0]},
    {[0.0, 0.0, 1.0], [0.0, 1.0, 1.0]}
  ]
  
rmse = Statistics.rootMeanSquareError(expectedWithOutputValues)