Library for creating AI with Dart

Created under a MIT-style license.

Overview

This library represent an simple way to create neural network.

There are 2 types of neural network that can be created:

  • MLP (multilayer and single-layer perceptron).
  • AE (autoencoder).

API

MLP

Perseptron is designed due to Rosenblatt's perseptron.

The main class is the MLP which can contains Layers. Each layer consist of one or many Neurons. First layer always must consist of input neurons where each of them take one input value and have weight equal to 1. All neurons of previous layer have contacts with each neurons of next layer.

In learning is used backpropagation algorithm.

AE

Architecture of AE is the same as MLP, except that first is used for encoding data.

Memory

Neural network have long-time and short-time memory. All information (knowledge - weights of synapses) of neural network during studying pass through short-time memory. When studying finished and knowledge is structured, then it pass to long-time memory. Knowledge is saved in JSON file knowledge.json in resources directory in the root of your library. For next time network take knowledge from file and initialize with proper weights.

Structure

Neural network can be created from predefined structure defined in structure.json file. You can place it anywhere you want, but default and preferred way is placing it in resources directory in the root of your library.

Every structure.json must have type property that correspond to neural network's name. Other properties such as:

- activation: specifies which function are used for neuron activation (default sigmoid)
- momentum: describe step of gradient descent (default 1)
- bias: is limit of neuron's choice. Used only in `step` function
- hyperparameter: is used in `PReLU`, `RReLU` and `ELU`. For `RReLU` it is a random number sampled from a uniform distribution `𝑈(𝑙, u)`, for `PReLU` it is a random value and for `ELU` it is random value that is equal or greater than zero

are optional. Specific properties for each structure type should be provided, except there are optional ones.

Structure of MLP:

{
  "type": "MLP",
  "activation": "relu",
  "input": 15,
  "hiddens": [3],
  "output": 3
}

Where input - count of input neurons, hiddens - array length shows count of hidden Layers and values are count of Neurons of each layer, output - count of output Neurons.

Structure of AE:

{
  "type": "AE",
  "input": 15,
  "hiddens": [3],
  "encoded": 3
}

Where input - count of input neurons, hiddens - array length shows count of hidden Layers and values are count of Neurons of each layer for encoded and decoded parts, decoded - count of decoded data Neurons.

Activation functions

Supported 9 activation functions:

1. step
2. sigmoid (default)
3. tanh
4. relu
5. srelu (smooth relu or softplus)
6. lrelu (leaky relu)
7. prelu (parametric relu)
8. rrelu (randomized relu)
9. elu (exponential linear units)

Visualization

Visualization of process of training network is available. Implemented only Mean Squared Error (MSE). If visualize parameter of train() method is true then MSE sends to console.

Sample

Available here.

Features and bugs

Please file feature requests and bugs at the issue tracker.

With ❤️ to AI

Libraries

ai
Library for creating AI with Dart [...]