Hopfield Neural Network in perl6

[holyghost]

A small implementation of an Artificial Neural Network using Hopfield neurons, synapses and a simple training system :

unit module ann;

use ann::HopfieldSynaps;

class HopfieldNeuron is export {

        has @.inputsynapses;
        has @.outputsynapses;

        has $.input;

        method BUILD($y1 = 1000000.rand) {
                $.input = $y1;

        }

        method fire() {

                ### with training update weights
                loop (my $i = 0; $i < @.inputsynapses.length; $i++) {
                        if (@.inputsynapses[$i].weight * @.inputsynaps
+es[$i].outputneuron.input >= 0) {
                                @.inputsynapses[$i].outputneuron.input
+ = 1;
                        } else {
                                @.inputsynapses[$i].outputneuron.input
+ = 0;
                        }
                }
        }


}

unit module ann;

use ann::HopfieldNeuron;

class HopfieldSynaps is export {

        has $.weight;
        has $.inputneuron;
        has $.outputneuron;

        method BUILD($inputneuron, $outputneuron, $y1 = 1000000.rand) 
+{
                $.weight = $y1;
        }

};

unit module ann;

use ann::HopfieldNeuron;
use ann::HopfieldSynaps;

class HopfieldNN is export {

        has @.neurons;

        method BUILD($size) {

                @.neurons = ();
                loop (my $n = 0; $n < $size; $n++) {

                        push (@.neurons, HopfieldNeuron.new());

                }

                loop (my $m = 0; $m < $size; $m++) {
                        loop (my $j = 0; $j < $size; $j++) {
                                push(@.neurons[$j].inputsynapses, Hopf
+ieldSynaps.new());
                                @.neurons[$j].inputsynapses[$j].output
+neuron =
                                        @.neurons[$m];

                        }
                }
                loop (my $i = 0; $i < $size; $i++) {
                        loop (my $j = 0; $j < $size; $j++) {
                                push(@.neurons[$j].outputsynapses, Hop
+fieldSynaps.new());
                                @.neurons[$j].outputsynapses[$j].outpu
+tneuron =
                                        @.neurons[$i];

                        }
                }

        }

        ### repeat this to train the network
        method start(@inputs) {

                ### the inputs length is less than the full neuron lis
+t
                ### the first neurons made in the constructor are the 
+inputs
                ### of the network

                loop (my $i = 0; $i < @inputs.length; $i++) {

                        @.neurons[$i].input = @inputs[$i];
                }
                loop (my $j = 0; $j < @.neurons.length; $j++) {

                        @.neurons[$j].fire();

                }
        }

        method start2(@inputs) {
                ### without any traning, first neurons are for the inp
+ut pattern

                loop (my $n = 0; $n < @inputs.length; $n++) {
                        @.neurons[$n].input = @inputs[$n];
                }

                loop (my $i = 0; $i < @.neurons.length; $i++) {
                        loop (my $j = 0; $j < @.neurons.length; $j++) 
+{
                                loop (my $k = 0; $k < @.neurons.length
+; $k++) {

                                        if ($k == $j) { next; };

                                @.neurons[$i].inputsynapses[$j].weight
+ += (2 * @.neurons[$i].inputsynapses[$j].outputneuron.input - 1) * (2
+ * @.neurons[$i].inputsynapes[$k].outputneuron.input -1);
                                }
                        }
                }
        }
};

[liz]

This looks pretty cool!

If you allow me to make some remarks: it looks like a very C-ish/Perl 5-ish piece of Perl 6 code.

And I think there is one error in it:

        method BUILD($inputneuron, $outputneuron, $y1 = 1000000.rand) 
+{
                $.weight = $y1;
        }

The $.weight is equivalent to self.weight. You can only use the accessor as a mutator if the attribute is marked is rw. So either you can change the assignment to $!weight = $y1 (directly accessing the attribute), or you can add is rw to the attribute declaration: has $.weight is rw. The choice is really whether you want the weight to be assignable from the outside or not and/or you want subclasses to be able to define their own "weight" method or not.

With regards to C-ism / Perl 5-isms: it feels to me that the following loop

loop (my $i = 0; $i < @.inputsynapses.length; $i++) {
    if (@.inputsynapses[$i].weight * @.inputsynapses[$i].outputneuron.
+input >= 0) {
        @.inputsynapses[$i].outputneuron.input = 1;
    } else {
        @.inputsynapses[$i].outputneuron.input = 0;
    }   
}

could be simplified to:

.outputneuron.input = +(.weight * .outputneuron.input >= 0)
  for @.inputsynapses;

This will iterate over all of the inputsynapses without needing to index. The +( ) is what transforms a Boolean True / False to 1 or 0.

Hope this helps