vxern / synadart

🧠 A simple, fully documented neural network library created for educational purposes, heavily inspired by the `ai` package.

synadart

The synadart library can be used to create neural networks of any complexity, as well as learn from the source code by studying its extremely clean implementation.

Launching our first network

To begin using the synadart, we must first import it into our project:

import 'package:synadart/synadart.dart';

Next, we must create a network of our chosen type. Let's create a sequential network, in which every layer has one input and one output tensor. This should be pretty easy:

final network = Sequential(learningRate: 0.3);

Our network is currently empty; it contains no layers and therefore no neurons. Let's add three layers; the input layer, one hidden layer and the output layer:

network.addLayers([
  Dense(15, activationAlgorithm: ActivationAlgorithm.sigmoid),
  Dense(5, activationAlgorithm: ActivationAlgorithm.sigmoid),
  Dense(1, activationAlgorithm: ActivationAlgorithm.sigmoid),
]);

Now that our network has some structure to it, we can begin using it.. No, not quite yet. Our network is still not trained, and has no clue what it is doing. Time to train it.

Firstly, we will create a list of expected values, i.e. values we are expecting the network to output. Here, we are expecting to get the number '5'.

final expected = [
  [0.01], // 0
  [0.01], // 1
  [0.01], // 2
  [0.01], // 3
  [0.01], // 4
  [0.99], // 5 ( This is what we are anticipating )
  [0.01], // 6
  [0.01], // 7
  [0.01], // 8
  [0.01], // 9
];

Fantastic, we are now expecting our infantile network to magically output a number 5, not having taught it a thing. Oh, right - that's where the training data part comes in!

We must now tell the network what each of our expected output values is associated with. Let's teach it some numbers:

final trainingData = [
  '111101101101111'.split('').map(double.parse).toList(), // Pixel representation of a 0,
  '001001001001001'.split('').map(double.parse).toList(), // a 1,
  '111001111100111'.split('').map(double.parse).toList(), // a 2,
  '111001111001111'.split('').map(double.parse).toList(), // a 3,
  '101101111001001'.split('').map(double.parse).toList(), // a 4,
  '111100111001111'.split('').map(double.parse).toList(), // a 5,
  '111100111101111'.split('').map(double.parse).toList(), // a 6,
  '111001001001001'.split('').map(double.parse).toList(), // a 7,
  '111101111101111'.split('').map(double.parse).toList(), // an 8,
  '111101111001111'.split('').map(double.parse).toList(), // and a 9.
];

Now that we granted our network a grand total of 10 numbers to learn, we can begin training the network using the values we've set up:

network.train(inputs: trainingData, expected: expected, iterations: 5000);

Wonderful! We've trained our network using the pixel representation of number images, and our network is now able to recognise the number '5' with relative confidence. The last step is to test our network's capabilities ourselves.

Let's give our network a couple pixel representations of distorted images of the number '5':

final testData = [
  '111100111000111'.split('').map(double.parse).toList(),
  '111100010001111'.split('').map(double.parse).toList(),
  '111100011001111'.split('').map(double.parse).toList(),
  '110100111001111'.split('').map(double.parse).toList(),
  '110100111001011'.split('').map(double.parse).toList(),
  '111100101001111'.split('').map(double.parse).toList(),
];

To check the confidence of the network in recognising distorted '5's:

for (final test in testData) {
  print('Confidence in recognising a distorted 5: ${network.process(test)}');
}

synadart's Issues

Relicense to allow usage of source code without disclosure.

The MIT or Apache licences would be more appropriate than GPL.

Save network?

Is there a way to save trained network and load it from?

There is no sense in Neural Network lib if the trained network can't be saved and loaded.