Newral

I recently started to learn about AI. Of course there are great libraries out there but I wanted to have something that makes it easy to test the different concepts to really understand them. Also I wanted to have a playground to easily see how good different approaches work for different data sets. I chose the name newral as its for newbies trying out neural networks and other AI related concepts

In the implementation I tried to write as little code as possible and used classes trying to avoid "array index hell". So the data structures are in no way tuned for efficiency, rather I tried to make clear what actually is going on. For every concept there should be at least one test to show it in action.

Install

gem install newral

What it does

Everything is still quite early stages but there are a lot of things you can do already

• Training Functions

  • Hill Climbing
  • Greedy
  • Gradient Descent

• K-Means Clustering

• K-Nearest Neighbour

• Neural Networks

  • Easily define simple ones often used in Tutorials
  • Backpropagation

• Graphs

  • Tree Search
  • Cheapest First
  • A Star

• Q-Learning Learn the computer to play Tic-Tac Toe (or other simple games )

I must say that this is really a total side project for me, so don´t expect lots of updates or bugfixes. Whenever I thought about it there are links to the tutorials or websites I used (which will explain the theory much better than I ever could). Please check out the tests where there are a few examples of possible use cases.

Stuff still in even earlier stages

• everything in genetic folder
• bayes / probability

So lets do some basic stuff

Error Calculation

lets assume we have 3 calculated results by our function and 3 expected outputs

current =  [1,2,3]
expected = [2,4,6]

so what´s the error

Newral::ErrorCalculation.mean_square( current, expected  )

same thing for vectors

 current = [
        [1,2,3],
        [3,9,16]

      ]

      expected = [
        [2,4,6],
        [4,8,9]
      ]

Newral::ErrorCalculation.mean_square( current, expected  )

Classifiers

points = [
  [1,1],[2,2],[4,4],
  [10,9],[11,12],[13,7]
].shuffle

n= Newral::Classifier::KMeansCluster.new( points, cluster_labels:[:cows,:elefants] ).process
n.clusters[:elefants].points
n.clusters[:cows].points

n=Newral::Classifier::Dendogram.new( points ).process
n.to_s

Neural Networks

create some neurons

perceptron = Newral::Networks::Perceptron.new(weights: [-2,-2],bias: 3) # look its a NAND gate
perceptron.update_with_vector [1,1]

sigmoid = Newral::Networks::Sigmoid.new(weights: [-2,-2],bias: 3) # sigmoids are much cooler 
sigmoid.update_with_vector [1,1]

create a basic network

      network = Newral::Networks::Network.define do 
        add_layer "input" do 
          add_neuron 'a', weights: [-2,-2],bias: 3, type: 'perceptron'
          add_neuron 'b', weights: [-2,-2],bias: 3, type: 'perceptron'
        end 
        add_layer "output" do 
          add_neuron 'c', weights: [-2,-2],bias: 3, type: 'perceptron'
        end 

        connect from:'a', to:'c'
        connect from:'b', to:'c'
      end

      network.update_with_vector [1,1]

create a network and perform backpropagation

inputs = [
        [0.05,0.1]
      ]
      outputs = [
        [0.01,0.99]
      ]
      network = Newral::Networks::BackpropagationNetwork.new( number_of_hidden: 2, number_of_outputs: 2)
      network.set_weights_and_bias( layer: 'hidden', weights:[[0.15,0.2],[0.25,0.3]],bias:[0.35,0.35])
      network.set_weights_and_bias( layer: 'output', weights:[[0.4,0.45],[0.5,0.55]], bias:[0.6,0.6])
      network.calculate_error( input: inputs, output: outputs ) # stupid network
      1000.times do 
        network.train input: inputs , output:outputs
      end 

      network.calculate_error( input: inputs, output: outputs ) # look it learned

Load some data

load the IRIS data set (Hello World of AI) located in test folder

data = Newral::Data::Csv.new(file_name:File.expand_path('../test/fixtures/IRIS.csv',__FILE__))
data.process
cluster_set = Newral::Classifier::KMeansCluster.new( data.inputs, cluster_labels: data.output_hash.keys ).process
cluster_set.clusters.length # There are 3 different types 

data = Newral::Data::Csv.new(file_name:File.expand_path('../test/fixtures/IRIS.csv',__FILE__))
data.process

network = Newral::Networks::BackpropagationNetwork.new( number_of_inputs: data.inputs.first.size, number_of_hidden: data.inputs.first.size, number_of_outputs: data.output_hash.keys.size )
network.calculate_error( input: data.inputs, output: data.output_as_vector ) # using a network with random weights  
100.times do
  network.train( input: data.inputs, output: data.output_as_vector ) # Hard training is the key to success in any neural nets life
end 
network.calculate_error( input: data.inputs, output: data.output_as_vector ) # hey it now knows flowers better than me!

Of course we don´t want oversampling so we should train and test on different data sets

data = Newral::Data::Csv.new(file_name:File.expand_path('../test/fixtures/IRIS.csv',__FILE__))
data.process

network = Newral::Networks::BackpropagationNetwork.new( number_of_inputs: data.inputs.first.size, number_of_hidden: data.inputs.first.size, number_of_outputs: data.output_hash.keys.size )
network.calculate_error( input: data.sub_set(set: :inputs, category: :validation ), output: data.output_as_vector( category: :validation ) ) 

100.times do
  network.train( input: data.sub_set(set: :inputs, category: :training ), output: data.output_as_vector( category: :training ) ) 
end 

network.calculate_error( input: data.sub_set(set: :inputs, category: :validation ), output: data.output_as_vector( category: :validation ) ) 

here comes the heavy stuff for this little library, load the MNIST data set (60000 images with 28*28 pixels). You can read more about MNIST http://yann.lecun.com/exdb/mnist/

data = Newral::Data::Idx.new( file_name:'~/Downloads/train-images-idx3-ubyte', label_file_name:'~/Downloads/train-labels-idx1-ubyte')
data.process

sample_data = data.sample( limit:100 ) 
sample_data.downsample_input!( width:2,height:2,width_of_line:28 ) # create less resolution pictures

sample_data2 = data.sample( limit:100, offset:100  ) # a 2nd sample
sample_data2.downsample_input!( width:2,height:2,width_of_line:28 )


network = Newral::Networks::BackpropagationNetwork.new( number_of_inputs: sample_data.inputs.first.size, number_of_hidden: sample_data.inputs.first.size, number_of_outputs: sample_data.output_hash.keys.size )

# lets compare the error of a random network vs one trained one
network.calculate_error( input: sample_data2.inputs, output: sample_data2.output_as_vector )

# use first sample to train
network.train( input: sample_data.inputs, output: sample_data.output_as_vector )

# now calculate the error of untrained sample
# it should still go down
network.calculate_error( input: sample_data2.inputs, output: sample_data2.output_as_vector )

use a tree Search to find the fastest path from Arad to Bucharest

edges,nodes,node_locations = setup_bulgarian_map # find this in the test folder 
      g = Newral::Graphs::Graph.new 
      g.add_nodes nodes
      g.add_edges edges
      t=Newral::Graphs::CheapestFirst.new( graph: g, start_node: 'Arad', end_node:'Bucharest')
      path = t.run
      path.cost

Use QLearning to play Tic Tac Toe

as we know good players will always reach a draw

require './test/games/tic_tac_toe_game'

  game = TicTacToeGame.new  # ( width: 8, height: 6, in_a_row: 4 )
      player1 = Newral::QLearning::Base.new( game: game, id: 0 )
      player2 = Newral::QLearning::Base.new( game: game, id: 1 )
      # training
      1000.times do
        game.run
        game.reset
      end
      game.reset( reset_score: 1 )
      player1.set_epsilon 1 # stop doing random moves, we know the game
      player2.set_epsilon 1

   game.run # => its a draw  

Use Training Algorithms to best approximate data with a function

Many typical functions suited for such approximations are already there

f= Newral::Functions::Vector.new vector: [1,6], bias:1 
f.calculate [4,7] #  4*1+6*7+1 => 47


Newral::Functions::Polynomial.new factors: [2,5,1] 
f.calculate 2 # 2*(2**2)+5*2+1 => 19

first lets use a basic polynominal function

      input = [2,4,8]
      output = [4,16,64] # best function is x**2, lets see if our training algorithms finds them
      g=Newral::Training::Greedy.new( input: input, output: output, klass: Newral::Functions::Polynomial )
      g.process 
      g.best_function.calculate_error( input: input, output: output )

      h=Newral::Training::HillClimbing.new( input: input, output: output, klass: Newral::Functions::Polynomial, start_function: g.best_function )
      h.process 
      h.best_function.calculate_error( input: input, output: output )

      # Gradient descent with error gradient approximation function 
      d=Newral::Training::GradientDescent.new( input: input, output: output, klass: Newral::Functions::Polynomial )
      d.process  
      d.best_function.calculate_error( input: input, output: output )

now lets use a Vector

      input = [[1,2],[2,4]]
      output=[3,7]
      g=Newral::Training::GradientDescent.new( input: input, output: output, klass: Newral::Functions::Vector )
      g.process
      g.best_function.calculate_error( input: input, output: output )