ai4r 1.12 → 2.0

data/lib/ai4r/neural_network/backpropagation.rb

@@ -1,162 +1,354 @@
+# frozen_string_literal: true
+
 # Author::    Sergio Fierens
 # License::   MPL 1.1
 # Project::   ai4r
-# Url::       http://ai4r.org/
+# Url::       https://github.com/SergioFierens/ai4r
 #
-# You can redistribute it and/or modify it under the terms of 
-# the Mozilla Public License version 1.1  as published by the 
+# You can redistribute it and/or modify it under the terms of
+# the Mozilla Public License version 1.1  as published by the
 # Mozilla Foundation at http://www.mozilla.org/MPL/MPL-1.1.txt
 
-require File.dirname(__FILE__) + '/../data/parameterizable' 
+require_relative '../data/parameterizable'
+require_relative 'activation_functions'
+require_relative 'weight_initializations'
 
 module Ai4r
-  
-  # Artificial Neural Networks are mathematical or computational models based on 
-  # biological neural networks. 
-  # 
+  # Artificial Neural Networks are mathematical or computational models based on
+  # biological neural networks.
+  #
   # More about neural networks:
-  # 
+  #
   # * http://en.wikipedia.org/wiki/Artificial_neural_network
   #
   module NeuralNetwork
-    
     # = Introduction
-    # 
+    #
     # This is an implementation of a multilayer perceptron network, using
     # the backpropagation algorithm for learning.
-    # 
-    # Backpropagation is a supervised learning technique (described 
-    # by Paul Werbos in 1974, and further developed by David E. 
+    #
+    # Backpropagation is a supervised learning technique (described
+    # by Paul Werbos in 1974, and further developed by David E.
     # Rumelhart, Geoffrey E. Hinton and Ronald J. Williams in 1986)
-    # 
+    #
     # = Features
-    # 
+    #
     # * Support for any network architecture (number of layers and neurons)
     # * Configurable propagation function
-    # * Optional usage of bias 
+    # * Optional usage of bias
     # * Configurable momentum
     # * Configurable learning rate
     # * Configurable initial weight function
     # * 100% ruby code, no external dependency
-    # 
+    #
     # = Parameters
-    # 
+    #
     # Use class method get_parameters_info to obtain details on the algorithm
     # parameters. Use set_parameters to set values for this parameters.
-    # 
-    # * :disable_bias => If true, the alforithm will not use bias nodes. 
+    #
+    # * :disable_bias => If true, the algorithm will not use bias nodes.
     #   False by default.
-    # * :initial_weight_function => f(n, i, j) must return the initial 
-    #   weight for the conection between the node i in layer n, and node j in 
+    # * :initial_weight_function => f(n, i, j) must return the initial
+    #   weight for the conection between the node i in layer n, and node j in
     #   layer n+1. By default a random number in [-1, 1) range.
-    # * :propagation_function => By default: 
+    # * :propagation_function => By default:
     #   lambda { |x| 1/(1+Math.exp(-1*(x))) }
-    # * :derivative_propagation_function => Derivative of the propagation 
-    #   function, based on propagation function output. 
+    # * :derivative_propagation_function => Derivative of the propagation
+    #   function, based on propagation function output.
     #   By default: lambda { |y| y*(1-y) }, where y=propagation_function(x)
-    # * :learning_rate => By default 0.25        
+    # * :activation => Built-in activation name (:sigmoid, :tanh or :relu).
+    #   Selecting this overrides propagation_function and derivative_propagation_function.
+    #   Default: :sigmoid
+    # * :learning_rate => By default 0.25
     # * :momentum => By default 0.1. Set this parameter to 0 to disable
     #   momentum
-    # 
+    #
     # = How to use it
-    # 
+    #
     #   # Create the network with 4 inputs, 1 hidden layer with 3 neurons,
     #   # and 2 outputs
-    #   net = Ai4r::NeuralNetwork::Backpropagation.new([4, 3, 2])  
+    #   net = Ai4r::NeuralNetwork::Backpropagation.new([4, 3, 2])
     #
-    #   # Train the network 
+    #   # Train the network
     #   1000.times do |i|
     #     net.train(example[i], result[i])
     #   end
-    #   
+    #
     #   # Use it: Evaluate data with the trained network
-    #   net.eval([12, 48, 12, 25])  
-    #     =>  [0.86, 0.01]   
-    #     
+    #   net.eval([12, 48, 12, 25])
+    #     =>  [0.86, 0.01]
+    #
     # More about multilayer perceptron neural networks and backpropagation:
-    # 
+    #
     # * http://en.wikipedia.org/wiki/Backpropagation
     # * http://en.wikipedia.org/wiki/Multilayer_perceptron
-    # 
+    #
     # = About the project
     # Author::    Sergio Fierens
     # License::   MPL 1.1
-    # Url::       http://ai4r.org
+    # Url::       https://github.com/SergioFierens/ai4r
     class Backpropagation
-      
       include Ai4r::Data::Parameterizable
-      
-      parameters_info :disable_bias => "If true, the alforithm will not use "+
-            "bias nodes. False by default.",
-        :initial_weight_function => "f(n, i, j) must return the initial "+
-            "weight for the conection between the node i in layer n, and "+
-            "node j in layer n+1. By default a random number in [-1, 1) range.",
-        :propagation_function => "By default: " + 
-            "lambda { |x| 1/(1+Math.exp(-1*(x))) }",         
-        :derivative_propagation_function => "Derivative of the propagation "+
-            "function, based on propagation function output. By default: " +
-            "lambda { |y| y*(1-y) }, where y=propagation_function(x)",
-        :learning_rate => "By default 0.25",        
-        :momentum => "By default 0.1. Set this parameter to 0 to disable "+
-            "momentum."
-          
+
       attr_accessor :structure, :weights, :activation_nodes, :last_changes
-      
+
+      # When the activation parameter changes, update internal lambdas for each
+      # layer. Accepts a single symbol or an array of symbols (one for each
+      # layer except the input layer).
+      # @param symbols [Object]
+      # @return [Object]
+      def activation=(symbols)
+        symbols = [symbols] unless symbols.is_a?(Array)
+        layer_count = @structure.length - 1
+        if symbols.length == 1
+          symbols = Array.new(layer_count, symbols.first)
+        elsif symbols.length != layer_count
+          raise ArgumentError, "Activation array size must match number of layers (#{layer_count})"
+        end
+        @activation = symbols
+        @propagation_functions = @activation.map do |a|
+          Ai4r::NeuralNetwork::ActivationFunctions::FUNCTIONS[a] ||
+            Ai4r::NeuralNetwork::ActivationFunctions::FUNCTIONS[:sigmoid]
+        end
+        @derivative_functions = @activation.map do |a|
+          Ai4r::NeuralNetwork::ActivationFunctions::DERIVATIVES[a] ||
+            Ai4r::NeuralNetwork::ActivationFunctions::DERIVATIVES[:sigmoid]
+        end
+      end
+
+      # @return [Object]
+      def activation
+        if @activation.is_a?(Array)
+          if @set_by_loss || (@loss_function == :cross_entropy && @activation_overridden)
+            @activation.first
+          else
+            @activation
+          end
+        else
+          @activation
+        end
+      end
+
+      # @param symbol [Object]
+      # @return [Object]
+      def weight_init=(symbol)
+        @weight_init = symbol
+        @initial_weight_function = case symbol
+                                   when :xavier
+                                     Ai4r::NeuralNetwork::WeightInitializations.xavier(@structure)
+                                   when :he
+                                     Ai4r::NeuralNetwork::WeightInitializations.he(@structure)
+                                   else
+                                     Ai4r::NeuralNetwork::WeightInitializations.uniform
+                                   end
+      end
+
+      # @param symbol [Object]
+      # @return [Object]
+      def loss_function=(symbol)
+        @loss_function = symbol
+        return unless symbol == :cross_entropy && !@activation_overridden && !@custom_propagation
+
+        @set_by_loss = true
+        self.activation = :softmax
+        @activation_overridden = false
+      end
+
       # Creates a new network specifying the its architecture.
       # E.g.
-      #    
+      #
       #   net = Backpropagation.new([4, 3, 2])  # 4 inputs
-      #                                         # 1 hidden layer with 3 neurons, 
-      #                                         # 2 outputs    
+      #                                         # 1 hidden layer with 3 neurons,
+      #                                         # 2 outputs
       #   net = Backpropagation.new([2, 3, 3, 4])   # 2 inputs
-      #                                             # 2 hidden layer with 3 neurons each, 
-      #                                             # 4 outputs    
+      #                                             # 2 hidden layer with 3 neurons each,
+      #                                             # 4 outputs
       #   net = Backpropagation.new([2, 1])   # 2 inputs
       #                                       # No hidden layer
-      #                                       # 1 output      
-      def initialize(network_structure)
+      #                                       # 1 output
+      # @param network_structure [Object]
+      # @param activation [Object]
+      # @param weight_init [Object]
+      # @return [Object]
+      def initialize(network_structure, activation = :sigmoid, weight_init = :uniform)
         @structure = network_structure
-        @initial_weight_function = lambda { |n, i, j| ((rand 2000)/1000.0) - 1}
-        @propagation_function = lambda { |x| 1/(1+Math.exp(-1*(x))) } #lambda { |x| Math.tanh(x) }
-        @derivative_propagation_function = lambda { |y| y*(1-y) } #lambda { |y| 1.0 - y**2 }
+        self.weight_init = weight_init
+        @custom_propagation = false
+        @set_by_loss = true
+        self.activation = activation
+        @activation_overridden = (activation != :sigmoid)
+        @set_by_loss = false
         @disable_bias = false
         @learning_rate = 0.25
         @momentum = 0.1
+        @loss_function = :mse
       end
 
-      # Evaluates the input.
-      # E.g.
-      #     net = Backpropagation.new([4, 3, 2])
       #     net.eval([25, 32.3, 12.8, 1.5])
       #         # =>  [0.83, 0.03]
+      # @param input_values [Object]
+      # @return [Object]
       def eval(input_values)
         check_input_dimension(input_values.length)
-        init_network if !@weights
+        init_network unless @weights
         feedforward(input_values)
-        return @activation_nodes.last.clone
+        @activation_nodes.last.clone
+      end
+
+      # Evaluates the input and returns most active node
+      # E.g.
+      #     net = Backpropagation.new([4, 3, 2])
+      #     net.eval_result([25, 32.3, 12.8, 1.5])
+      #         # eval gives [0.83, 0.03]
+      #         # =>  0
+      # @param input_values [Object]
+      # @return [Object]
+      def eval_result(input_values)
+        result = eval(input_values)
+        result.index(result.max)
       end
-      
+
       # This method trains the network using the backpropagation algorithm.
-      # 
+      #
       # input: Networks input
-      # 
+      #
       # output: Expected output for the given input.
       #
-      # This method returns the network error:
-      # => 0.5 * sum( (expected_value[i] - output_value[i])**2 )
+      # This method returns the training loss according to +loss_function+.
+      # @param inputs [Object]
+      # @param outputs [Object]
+      # @return [Object]
       def train(inputs, outputs)
         eval(inputs)
         backpropagate(outputs)
-        calculate_error(outputs)
+        calculate_loss(outputs, @activation_nodes.last)
+      end
+
+      # Train a list of input/output pairs and return average loss.
+      # @param batch_inputs [Object]
+      # @param batch_outputs [Object]
+      # @return [Object]
+      def train_batch(batch_inputs, batch_outputs)
+        if batch_inputs.length != batch_outputs.length
+          raise ArgumentError,
+                'Inputs and outputs size mismatch'
+        end
+
+        batch_size = batch_inputs.length
+        init_network unless @weights
+
+        accumulated_changes = Array.new(@weights.length) do |w|
+          Array.new(@weights[w].length) do |i|
+            Array.new(@weights[w][i].length, 0.0)
+          end
+        end
+
+        sum_error = 0.0
+        batch_inputs.each_index do |idx|
+          inputs = batch_inputs[idx]
+          outputs = batch_outputs[idx]
+          eval(inputs)
+          calculate_output_deltas(outputs)
+          calculate_internal_deltas
+
+          (@weights.length - 1).downto(0) do |n|
+            @weights[n].each_index do |i|
+              @weights[n][i].each_index do |j|
+                change = @deltas[n][j] * @activation_nodes[n][i]
+                accumulated_changes[n][i][j] += change
+              end
+            end
+          end
+
+          sum_error += calculate_loss(outputs, @activation_nodes.last)
+        end
+
+        (@weights.length - 1).downto(0) do |n|
+          @weights[n].each_index do |i|
+            @weights[n][i].each_index do |j|
+              avg_change = accumulated_changes[n][i][j] / batch_size.to_f
+              @weights[n][i][j] += (learning_rate * avg_change) + (momentum * @last_changes[n][i][j])
+              @last_changes[n][i][j] = avg_change
+            end
+          end
+        end
+
+        sum_error / batch_size.to_f
       end
-      
-      # Initialize (or reset) activation nodes and weights, with the 
+
+      # Train for a number of epochs over the dataset. Optionally define a batch size.
+      # Data can be shuffled between epochs passing +shuffle: true+ (default).
+      # Use +random_seed+ to make shuffling deterministic.
+      # Returns an array with the average loss of each epoch.
+      # @return [Object]
+      def train_epochs(data_inputs, data_outputs, epochs:, batch_size: 1,
+                       early_stopping_patience: nil, min_delta: 0.0,
+                       shuffle: true, random_seed: nil, &block)
+        if data_inputs.length != data_outputs.length
+          raise ArgumentError,
+                'Inputs and outputs size mismatch'
+        end
+
+        losses = []
+        best_loss = Float::INFINITY
+        patience = early_stopping_patience
+        patience_counter = 0
+        rng = random_seed.nil? ? Random.new : Random.new(random_seed)
+        epochs.times do |epoch|
+          epoch_error = 0.0
+          epoch_inputs = data_inputs
+          epoch_outputs = data_outputs
+          if shuffle
+            indices = (0...data_inputs.length).to_a.shuffle(random: rng)
+            epoch_inputs = data_inputs.values_at(*indices)
+            epoch_outputs = data_outputs.values_at(*indices)
+          end
+          index = 0
+          while index < epoch_inputs.length
+            batch_in = epoch_inputs[index, batch_size]
+            batch_out = epoch_outputs[index, batch_size]
+            batch_error = train_batch(batch_in, batch_out)
+            epoch_error += batch_error * batch_in.length
+            index += batch_size
+          end
+          epoch_loss = epoch_error / data_inputs.length.to_f
+          losses << epoch_loss
+          if block
+            if block.arity >= 3
+              correct = 0
+              data_inputs.each_index do |i|
+                output = eval(data_inputs[i])
+                predicted = output.index(output.max)
+                expected = data_outputs[i].index(data_outputs[i].max)
+                correct += 1 if predicted == expected
+              end
+              accuracy = correct.to_f / data_inputs.length
+              block.call(epoch, epoch_loss, accuracy)
+            else
+              block.call(epoch, epoch_loss)
+            end
+          end
+          if patience
+            if best_loss - epoch_loss > min_delta
+              best_loss = epoch_loss
+              patience_counter = 0
+            else
+              patience_counter += 1
+              break if patience_counter >= patience
+            end
+          end
+        end
+        losses
+      end
+
+      # Initialize (or reset) activation nodes and weights, with the
       # provided net structure and parameters.
+      # @return [Object]
       def init_network
         init_activation_nodes
         init_weights
         init_last_changes
-        return self
+        self
       end
 
       protected
@@ -168,6 +360,7 @@ module Ai4r
       # * propagation_function
       # * derivative_propagation_function
       # you must restore their values manually after loading the instance.
+      # @return [Object]
       def marshal_dump
         [
           @structure,
@@ -176,151 +369,234 @@ module Ai4r
           @momentum,
           @weights,
           @last_changes,
-          @activation_nodes
+          @activation_nodes,
+          @activation
         ]
-     end
-
-     def marshal_load(ary)
-       @structure,
-          @disable_bias,
-          @learning_rate,
-          @momentum,
-          @weights,
-          @last_changes,
-          @activation_nodes = ary
-       @initial_weight_function = lambda { |n, i, j| ((rand 2000)/1000.0) - 1}
-       @propagation_function = lambda { |x| 1/(1+Math.exp(-1*(x))) } #lambda { |x| Math.tanh(x) }
-       @derivative_propagation_function = lambda { |y| y*(1-y) } #lambda { |y| 1.0 - y**2 }
-     end
+      end
 
+      # @param ary [Object]
+      # @return [Object]
+      def marshal_load(ary)
+        @structure,
+           @disable_bias,
+           @learning_rate,
+           @momentum,
+           @weights,
+           @last_changes,
+           @activation_nodes,
+           @activation = ary
+        self.weight_init = :uniform
+        self.activation = @activation || :sigmoid
+      end
 
       # Propagate error backwards
+      # @param expected_output_values [Object]
+      # @return [Object]
       def backpropagate(expected_output_values)
         check_output_dimension(expected_output_values.length)
         calculate_output_deltas(expected_output_values)
         calculate_internal_deltas
         update_weights
       end
-      
+
       # Propagate values forward
+      # @param input_values [Object]
+      # @return [Object]
       def feedforward(input_values)
-        input_values.each_index do |input_index| 
+        input_values.each_index do |input_index|
           @activation_nodes.first[input_index] = input_values[input_index]
         end
         @weights.each_index do |n|
-          @structure[n+1].times do |j|
-            sum = 0.0
+          sums = Array.new(@structure[n + 1], 0.0)
+          @structure[n + 1].times do |j|
             @activation_nodes[n].each_index do |i|
-              sum += (@activation_nodes[n][i] * @weights[n][i][j])
+              sums[j] += (@activation_nodes[n][i] * @weights[n][i][j])
+            end
+          end
+          if @activation[n] == :softmax
+            values = @propagation_functions[n].call(sums)
+            values.each_index { |j| @activation_nodes[n + 1][j] = values[j] }
+          else
+            sums.each_index do |j|
+              @activation_nodes[n + 1][j] = @propagation_functions[n].call(sums[j])
             end
-            @activation_nodes[n+1][j] = @propagation_function.call(sum)
           end
-        end        
+        end
       end
-      
+
       # Initialize neurons structure.
+      # @return [Object]
       def init_activation_nodes
-        @activation_nodes = Array.new(@structure.length) do |n| 
+        @activation_nodes = Array.new(@structure.length) do |n|
           Array.new(@structure[n], 1.0)
         end
-        if not disable_bias
-          @activation_nodes[0...-1].each {|layer| layer << 1.0 }
-        end
+        return if disable_bias
+
+        @activation_nodes[0...-1].each { |layer| layer << 1.0 }
       end
-      
+
       # Initialize the weight arrays using function specified with the
       # initial_weight_function parameter
+      # @return [Object]
       def init_weights
-        @weights = Array.new(@structure.length-1) do |i|
+        @weights = Array.new(@structure.length - 1) do |i|
           nodes_origin = @activation_nodes[i].length
-          nodes_target = @structure[i+1]
+          nodes_target = @structure[i + 1]
           Array.new(nodes_origin) do |j|
-            Array.new(nodes_target) do |k| 
+            Array.new(nodes_target) do |k|
               @initial_weight_function.call(i, j, k)
             end
           end
         end
-      end   
+      end
 
-      # Momentum usage need to know how much a weight changed in the 
-      # previous training. This method initialize the @last_changes 
+      # Momentum usage need to know how much a weight changed in the
+      # previous training. This method initialize the @last_changes
       # structure with 0 values.
+      # @return [Object]
       def init_last_changes
         @last_changes = Array.new(@weights.length) do |w|
-          Array.new(@weights[w].length) do |i| 
+          Array.new(@weights[w].length) do |i|
             Array.new(@weights[w][i].length, 0.0)
           end
         end
       end
-      
+
       # Calculate deltas for output layer
+      # @param expected_values [Object]
+      # @return [Object]
       def calculate_output_deltas(expected_values)
         output_values = @activation_nodes.last
         output_deltas = []
+        func = @derivative_functions.last
         output_values.each_index do |output_index|
-          error = expected_values[output_index] - output_values[output_index]
-          output_deltas << @derivative_propagation_function.call(
-            output_values[output_index]) * error
+          if @loss_function == :cross_entropy && @activation == :softmax
+            output_deltas << (output_values[output_index] - expected_values[output_index])
+          else
+            error = expected_values[output_index] - output_values[output_index]
+            output_deltas << (func.call(output_values[output_index]) * error)
+          end
         end
         @deltas = [output_deltas]
       end
-      
+
       # Calculate deltas for hidden layers
+      # @return [Object]
       def calculate_internal_deltas
         prev_deltas = @deltas.last
-        (@activation_nodes.length-2).downto(1) do |layer_index|
+        (@activation_nodes.length - 2).downto(1) do |layer_index|
           layer_deltas = []
           @activation_nodes[layer_index].each_index do |j|
             error = 0.0
-            @structure[layer_index+1].times do |k|
+            @structure[layer_index + 1].times do |k|
               error += prev_deltas[k] * @weights[layer_index][j][k]
             end
-            layer_deltas[j] = (@derivative_propagation_function.call(
-              @activation_nodes[layer_index][j]) * error)
+            func = @derivative_functions[layer_index - 1]
+            layer_deltas[j] = func.call(@activation_nodes[layer_index][j]) * error
           end
           prev_deltas = layer_deltas
           @deltas.unshift(layer_deltas)
         end
       end
-      
+
       # Update weights after @deltas have been calculated.
+      # @return [Object]
       def update_weights
-        (@weights.length-1).downto(0) do |n|
-          @weights[n].each_index do |i|  
-            @weights[n][i].each_index do |j|  
-              change = @deltas[n][j]*@activation_nodes[n][i]
-              @weights[n][i][j] += ( learning_rate * change + 
-                  momentum * @last_changes[n][i][j])
+        (@weights.length - 1).downto(0) do |n|
+          @weights[n].each_index do |i|
+            @weights[n][i].each_index do |j|
+              change = @deltas[n][j] * @activation_nodes[n][i]
+              @weights[n][i][j] += ((learning_rate * change) +
+                  (momentum * @last_changes[n][i][j]))
               @last_changes[n][i][j] = change
             end
           end
         end
       end
-      
-      # Calculate quadratic error for a expected output value 
+
+      # Calculate quadratic error for an expected output value
       # Error = 0.5 * sum( (expected_value[i] - output_value[i])**2 )
+      # @param expected_output [Object]
+      # @return [Object]
       def calculate_error(expected_output)
         output_values = @activation_nodes.last
         error = 0.0
         expected_output.each_index do |output_index|
-          error += 
-            0.5*(output_values[output_index]-expected_output[output_index])**2
+          error +=
+            0.5 * ((output_values[output_index] - expected_output[output_index])**2)
         end
-        return error
+        error
       end
-      
+
+      # Calculate loss for expected/actual vectors according to selected
+      # loss_function (:mse or :cross_entropy).
+      # @param expected [Object]
+      # @param actual [Object]
+      # @return [Object]
+      def calculate_loss(expected, actual)
+        case @loss_function
+        when :cross_entropy
+          epsilon = 1e-12
+          loss = 0.0
+          if @activation == :softmax
+            expected.each_index do |i|
+              p = [[actual[i], epsilon].max, 1 - epsilon].min
+              loss -= expected[i] * Math.log(p)
+            end
+          else
+            expected.each_index do |i|
+              p = [[actual[i], epsilon].max, 1 - epsilon].min
+              loss -= (expected[i] * Math.log(p)) + ((1 - expected[i]) * Math.log(1 - p))
+            end
+          end
+          loss
+        else
+          # Mean squared error
+          error = 0.0
+          expected.each_index do |i|
+            error += 0.5 * ((expected[i] - actual[i])**2)
+          end
+          error
+        end
+      end
+
+      # @param inputs [Object]
+      # @return [Object]
       def check_input_dimension(inputs)
-        raise ArgumentError, "Wrong number of inputs. " +
-          "Expected: #{@structure.first}, " +
-          "received: #{inputs}." if inputs!=@structure.first
+        return unless inputs != @structure.first
+
+        raise ArgumentError, 'Wrong number of inputs. ' \
+                             "Expected: #{@structure.first}, " \
+                             "received: #{inputs}."
       end
 
+      # @param outputs [Object]
+      # @return [Object]
       def check_output_dimension(outputs)
-        raise ArgumentError, "Wrong number of outputs. " +
-          "Expected: #{@structure.last}, " +
-          "received: #{outputs}." if outputs!=@structure.last
+        return unless outputs != @structure.last
+
+        raise ArgumentError, 'Wrong number of outputs. ' \
+                             "Expected: #{@structure.last}, " \
+                             "received: #{outputs}."
       end
-      
+
+      parameters_info disable_bias: 'If true, the algorithm will not use ' \
+                                   'bias nodes. False by default.',
+                      initial_weight_function: 'f(n, i, j) must return the initial ' \
+                                               'weight for the conection between the node i in layer n, and ' \
+                                               'node j in layer n+1. By default a random number in [-1, 1) range.',
+                      weight_init: 'Built-in weight initialization strategy (:uniform, :xavier or :he). Default: :uniform',
+                      propagation_function: 'By default: ' \
+                                            'lambda { |x| 1/(1+Math.exp(-1*(x))) }',
+                      derivative_propagation_function: 'Derivative of the propagation ' \
+                                                       'function, based on propagation function output. By default: ' \
+                                                       'lambda { |y| y*(1-y) }, where y=propagation_function(x)',
+                      activation: 'Activation function per layer. Provide a symbol or an array of symbols (:sigmoid, :tanh, :relu or :softmax). Default: :sigmoid',
+                      learning_rate: 'By default 0.25',
+                      momentum: 'By default 0.1. Set this parameter to 0 to disable ' \
+                                'momentum.',
+                      loss_function: 'Loss function used when training (:mse or ' \
+                                     ':cross_entropy). Default: :mse'
     end
   end
 end