A Neural Network prototype in Elixir (Under development)
Used as part of my Transcendence in Erlang Talk, to demonstrate how Erlang processes can represent Neurons in an elegant way.
All Neurons (nodes) at present are GenServers with SimpleNeuron struct state.
- Input Nodes: detected as they do not have any registered in connections therefore forward sensor input.
- Hidden Nodes: use hyperbolic tangent activation function + wait for 1 input from each in node before becoming activated.
- Output Nodes (not implemented): can be detected the same way as input neurons are, but will apply Softmax activation function.
Simple Feed Forward Demo:
# Create our neurons
{:ok, n1} = SimpleNeuron.start_link()
{:ok, n2} = SimpleNeuron.start_link()
{:ok, n3} = SimpleNeuron.start_link()
{:ok, n4} = SimpleNeuron.start_link([bias: 0.13])
# Create the synaptic connections
SimpleNeuron.connect(n1, n4, 0.01)
SimpleNeuron.connect(n2, n4, 0.05)
SimpleNeuron.connect(n3, n4, 0.09)
# Trigger input neurons
SimpleNeuron.signal(n1, 1)
SimpleNeuron.signal(n2, 2)
SimpleNeuron.signal(n3, 3)$ iex -S mix
ex(1)> Nnex.example
13:41:07.418 [debug] Creating node with bias: 0.13
13:41:07.418 [info] <0.91.0> Received signal 1 from <0.89.0> input sensor
13:41:07.421 [info] <0.92.0> Received signal 2 from <0.89.0> input sensor
13:41:07.426 [info] <0.93.0> Received signal 3 from <0.89.0> input sensor
13:41:07.426 [info] <0.91.0> sending value 1 to [<0.94.0>]
13:41:07.430 [info] <0.92.0> sending value 2 to [<0.94.0>]
13:41:07.430 [info] <0.93.0> sending value 3 to [<0.94.0>]
13:41:07.430 [info] <0.94.0> Received signal 1 from <0.91.0> with connection weight 0.01
13:41:07.430 [info] <0.94.0> Received signal 2 from <0.92.0> with connection weight 0.05
13:41:07.430 [info] <0.94.0> Received signal 3 from <0.93.0> with connection weight 0.09
13:41:07.430 [info] <0.94.0> ACTIVATED!
13:41:07.430 [info] Apply weight: 0.09 to value: 3
13:41:07.430 [info] Apply weight: 0.05 to value: 2
13:41:07.430 [info] Apply weight: 0.01 to value: 1
13:41:07.430 [info] Apply weight: 1 to value: 0.13
13:41:07.431 [info] <0.94.0> sending value 0.46994519893303766 to []0.46994519893303766 represents the value node 4 computed from 3 input nodes.