NeuronSystem

The very simple example of the implementation of the neuron network on Elixir.

Installation

If available in Hex, the package can be installed as:

1. Add neuron_system to your list of dependencies in mix.exs:

```elixir
def deps do
  [{:neuron_system, "~> 0.1.0"}]
end
```

2. Ensure neuron_system is started before your application:

```elixir
def application do
  [applications: [:neuron_system]]
end
```

Usage

Net creation

The main model is Net. When you start to build your own net the first step is to create new one:

net = NeuronSystem.Net.create
# => %NeuronSystem.Models.Net{pid: #PID<0.198.0>}

Neuron adding

The next action is to add neurons to the Net:

{:ok, neuron} = NeuronSystem.Net.add_neuron(net, fn(value) -> 1 end)
{:ok,
 %NeuronSystem.Models.Neuron{activation_function: #Function<6.54118792/1 in :erl_eval.expr/5>,
  id: "neuron:a1d2730c-4ffc-11e6-bf11-a820663dc092"}}

The first parameter is the Net into which you add a new neuron. The second parameter is an activation function.

Connections

There are several types of connections:

1. Input connection
2. Connection between neurons
3. Output connection

Let consider each type.

Input connection

By adding input connection you define an input for the Net. Please, consider an example below:

NeuronSystem.Net.add_connection(net, :in, neuron, 0.3, :x1)

We've added an input connection for the given net, which is connected to the given neuron. That connection has a some weight (0.3) and it has a name or key. This key will be used later, when you do activation. Income map must has an entry with the previously mentioned key. Otherwise your net will stuck.

Connection between neurons

It is very easy and pretty straitforward:

NeuronSystem.Net.add_connection(net, neuron1, neuron2, 0.2)

We've added a connection between neuron1 and neuron2 with a weight 0.2.

Output connection

Output connection is used to define an output of the Net. For example:

NeuronSystem.Net.add_connection(net, :out, neuron1, 1.0, :y1)

In the example above, we've specified an output connection which is named y1. The source for the output is neuron1. And, as like as other connections, it has a weight (1.0).

Activation

Once you've built your Net you can to activate it! This is the main reason to build it, right?) After all preparation you can do it by the following code:

income = %{x1: 1}
NeuronSystem.Net.activate!(net, income)

Of course, you must specify an income map. And after some time you will see a result, for example:

# => [y1: 0]

Yay!

More difficult examples

Example of a Net with single neuron

income = %{x1: 0.5}
net = NeuronSystem.Net.create
{:ok, neuron1} = NeuronSystem.Net.add_neuron(net, fn
  x when x >= 1 -> 1
  x when x < 1 -> 0
end)
NeuronSystem.Net.add_connection(net, :in, neuron1, 1.0, :x1)
NeuronSystem.Net.add_connection(net, :out, neuron1, 1.0, :y1)
NeuronSystem.Net.activate!(net, income)
# => [y1: 0]

Example with two neurons in a line

income = %{x1: 3}
net = NeuronSystem.Net.create
{:ok, neuron1} = NeuronSystem.Net.add_neuron(net, fn
  x when x >= 0.5 -> 1
  x when x < 0.5 -> 0
end)
{:ok, neuron2} = NeuronSystem.Net.add_neuron(net, fn
  x when x >= 0.2 -> 1
  x when x < 0.2 -> 0
end)
NeuronSystem.Net.add_connection(net, :in, neuron1, 0.3, :x1)
NeuronSystem.Net.add_connection(net, neuron1, neuron2, 0.2)
NeuronSystem.Net.add_connection(net, :out, neuron2, 1.0, :y1)
NeuronSystem.Net.activate!(net, income)
# => [y1: 1]

Further ideas

1. The main next goal - implement learning algorithm
2. Add ability to create different types of the neurons using the high-level API