Module: Ai4r::NeuralNetwork::ActivationFunctions

Defined in:

lib/ai4r/neural_network/activation_functions.rb

Overview

Collection of common activation functions and their derivatives.

Constant Summary

FUNCTIONS =

{
  sigmoid: ->(x) { 1.0 / (1.0 + Math.exp(-x)) },
  tanh: ->(x) { Math.tanh(x) },
  relu: ->(x) { [x, 0].max },

  softmax: lambda do |arr|
    max = arr.max
    exps = arr.map { |v| Math.exp(v - max) }
    sum = exps.inject(:+)
    exps.map { |e| e / sum }
  end
}.freeze

DERIVATIVES =

{
  sigmoid: ->(y) { y * (1 - y) },
  tanh: ->(y) { 1.0 - (y**2) },
  relu: ->(y) { y.positive? ? 1.0 : 0.0 },
  softmax: ->(y) { y * (1 - y) }
}.freeze