AI-NeuralNet-Mesh
view release on metacpan or search on metacpan
low()
:acts
- Exports:
ramp()
and_gate()
or_gate()
</PRE>
<P>See the respective methods/functions for information about
each method/functions usage.</P>
<P>
<HR>
<H1><A NAME="methods">METHODS</A></H1>
<DL>
<DT><STRONG><A NAME="item_new">AI::NeuralNet::Mesh->new();</A></STRONG><BR>
<DD>
There are four ways to construct a new network with new(). Each is detailed below.
<P>P.S. Don't worry, the old <A HREF="#item_new"><CODE>new($layers, $nodes [, $outputs])</CODE></A> still works like always!</P>
<P></P>
<DT><STRONG>AI::NeuralNet::Mesh->new($layers, $nodes [, $outputs]);</STRONG><BR>
<DD>
Returns a newly created neural network from an <CODE>AI::NeuralNet::Mesh</CODE>
object. The network will have <CODE>$layers</CODE> number of layers in it
and it will have <CODE>$nodes</CODE> number of nodes per layer.
<P>There is an optional parameter of $outputs, which specifies the number
of output neurons to provide. If $outputs is not specified, $outputs
defaults to equal $size.</P>
<P></P>
<DT><STRONG>AI::NeuralNet::Mesh->new($file);</STRONG><BR>
<DD>
This will automatically create a new network from the file <CODE>$file</CODE>. It will
return undef if the file was of an incorrect format or non-existant. Otherwise,
it will return a blessed refrence to a network completly restored from <CODE>$file</CODE>.
<P></P>
<DT><STRONG>AI::NeuralNet::Mesh->new(\@layer_sizes);</STRONG><BR>
<DD>
This constructor will make a network with the number of layers corresponding to the length
in elements of the array ref passed. Each element in the array ref passed is expected
to contain an integer specifying the number of nodes (neurons) in that layer. The first
layer ($layer_sizes[0]) is to be the input layer, and the last layer in @layer_sizes is to be
the output layer.
<P>Example:</P>
<PRE>
my $net = AI::NeuralNet::Mesh->new([2,3,1]);</PRE>
<P>Creates a network with 2 input nodes, 3 hidden nodes, and 1 output node.</P>
<P></P>
<DT><STRONG>AI::NeuralNet::Mesh->new(\@array_of_hashes);</STRONG><BR>
<DD>
Another dandy constructor...this is my favorite. It allows you to tailor the number of layers,
the size of the layers, the activation type (you can even add anonymous inline subs with this one),
and even the threshold, all with one array ref-ed constructor.
<P>Example:</P>
<PRE>
my $net = AI::NeuralNet::Mesh->new([
{
nodes => 2,
activation => linear
},
{
nodes => 3,
activation => sub {
my $sum = shift;
return $sum + rand()*1;
}
},
{
nodes => 1,
activation => sigmoid,
threshold => 0.75
}
]);
</PRE>
<P>Interesting, eh? What you are basically passing is this:</P>
<PRE>
my @info = (
{ },
{ },
{ },
...
);</PRE>
<P>You are passing an array ref who's each element is a hash refrence. Each
hash refrence, or more precisely, each element in the array refrence you are passing
to the constructor, represents a layer in the network. Like the constructor above,
the first element is the input layer, and the last is the output layer. The rest are
hidden layers.</P>
<P>Each hash refrence is expected to have AT LEAST the ``nodes'' key set to the number
of nodes (neurons) in that layer. The other two keys are optional. If ``activation'' is left
out, it defaults to ``linear''. If ``threshold'' is left out, it defaults to 0.50.</P>
<P>The ``activation'' key can be one of four values:</P>
<PRE>
linear ( simply use sum of inputs as output )
sigmoid [ sigmoid_1 ] ( only positive sigmoid )
sigmoid_2 ( positive / 0 /negative sigmoid )
\&code_ref;</PRE>
<P>``sigmoid_1'' is an alias for ``sigmoid''.</P>
<P>The code ref option allows you to have a custom activation function for that layer.
The code ref is called with this syntax:</P>
<PRE>
$output = &$code_ref($sum_of_inputs, $self);
</PRE>
<P>The code ref is expected to return a value to be used as the output of the node.
The code ref also has access to all the data of that node through the second argument,
a blessed hash refrence to that node.</P>
<P>See CUSTOM ACTIVATION FUNCTIONS for information on several included activation functions
other than the ones listed above.</P>
<P>Three of the activation syntaxes are shown in the first constructor above, the ``linear'',
``sigmoid'' and code ref types.</P>
<P>You can also set the activation and threshold values after network creation with the
<A HREF="#item_activation"><CODE>activation()</CODE></A> and <A HREF="#item_threshold"><CODE>threshold()</CODE></A> methods.</P>
<P></P>
<P></P>
<DT><STRONG><A NAME="item_learn">$net->learn($input_map_ref, $desired_result_ref [, options ]);</A></STRONG><BR>
<DD>
NOTE: <A HREF="#item_learn_set"><CODE>learn_set()</CODE></A> now has increment-degrading turned OFF by default. See note
on the degrade flag, below.
<P>This will 'teach' a network to associate an new input map with a desired
result. It will return a string containg benchmarking information.</P>
<P>You can also specify strings as inputs and ouputs to learn, and they will be
crunched automatically. Example:</P>
<PRE>
$net->learn('corn', 'cob');
occured.
<P></P>
<DT><STRONG><A NAME="item_load_pcx">$net->load_pcx($filename);</A></STRONG><BR>
<DD>
NOTE: To use this function, you must have PCX::Loader installed. If you do not have
PCX::Loader installed, it will return undef and store an error for you to retrive with
the <A HREF="#item_error"><CODE>error()</CODE></A> method, below.
<P>This is a treat... this routine will load a PCX-format file (yah, I know ... ancient
format ... but it is the only one I could find specs for to write it in Perl. If
anyone can get specs for any other formats, or could write a loader for them, I
would be very grateful!) Anyways, a PCX-format file that is exactly 320x200 with 8 bits
per pixel, with pure Perl. It returns a blessed refrence to a PCX::Loader object, which
supports the following routinges/members. See example files ex_pcx.pl and ex_pcxl.pl in
the ./examples/ directory.</P>
<P>See <CODE>perldoc PCX::Loader</CODE> for information on the methods of the object returned.</P>
<P>You can download PCX::Loader from <A HREF="http://www.josiah.countystart.com/modules/get.pl?pcx-loader:mpod">http://www.josiah.countystart.com/modules/get.pl?pcx-loader:mpod</A></P>
<P></P></DL>
<P>
<HR>
<H1><A NAME="custom activation functions">CUSTOM ACTIVATION FUNCTIONS</A></H1>
<P>Included in this package are four custom activation functions meant to be used
as a guide to create your own, as well as to be useful to you in normal use of the
module. There is only one function exported by default into your namespace, which
is the <A HREF="#item_range"><CODE>range()</CODE></A> functions. These are not meant to be used as methods, but as functions.
These functions return code refs to a Perl closure which does the actual work when
the time comes.</P>
<DL>
<DT><STRONG>range(0..X);</STRONG><BR>
<DD>
<DT><STRONG>range(@range);</STRONG><BR>
<DD>
<DT><STRONG>range(A,B,C);</STRONG><BR>
<DD>
<A HREF="#item_range"><CODE>range()</CODE></A> returns a closure limiting the output
of that node to a specified set of values.
Good for use in output layers.
<P>Usage example:
$net->activation(4,range(0..5));
or (in the <A HREF="#item_new"><CODE>new()</CODE></A> hash constructor form):
..
{
nodes => 1,
activation => range 5..2
}
..
You can also pass an array containing the range
values (not array ref), or you can pass a comma-
seperated list of values as parameters:</P>
<PRE>
$net->activation(4,range(@numbers));
$net->activation(4,range(6,15,26,106,28,3));</PRE>
<P>Note: when using a <A HREF="#item_range"><CODE>range()</CODE></A> activatior, train the
net TWICE on the data set, because the first time
the <A HREF="#item_range"><CODE>range()</CODE></A> function searches for the top value in
the inputs, and therefore, results could flucuate.
The second learning cycle guarantees more accuracy.</P>
<P>The actual code that implements the range closure is
a bit convulted, so I will expand on it here as a simple
tutorial for custom activation functions.</P>
<PRE>
= line 1 = sub {
= line 2 = my @values = ( 6..10 );
= line 3 = my $sum = shift;
= line 4 = my $self = shift;
= line 5 = $self->{top_value}=$sum if($sum>$self->{top_value});
= line 6 = my $index = intr($sum/$self->{top_value}*$#values);
= line 7 = return $values[$index];
= line 8 = }</PRE>
<P>Now, the actual function fits in one line of code, but I expanded it a bit
here. Line 1 creates our array of allowed output values. Lines two and
three grab our parameters off the stack which allow us access to the
internals of this node. Line 5 checks to see if the sum output of this
node is higher than any previously encountered, and, if so, it sets
the marker higher. This also shows that you can use the $self refrence
to maintain information across activations. This technique is also used
in the <A HREF="#item_ramp"><CODE>ramp()</CODE></A> activator. Line 6 computes the index into the allowed
values array by first scaling the $sum to be between 0 and 1 and then
expanding it to fit smoothly inside the number of elements in the array. Then
we simply round to an integer and pluck that index from the array and
use it as the output value for that node.</P>
<P>See? It's not that hard! Using custom activation functions, you could do
just about anything with the node that you want to, since you have
access to the node just as if you were a blessed member of that node's object.</P>
<P></P>
<DT><STRONG><A NAME="item_ramp">ramp($r);</A></STRONG><BR>
<DD>
<A HREF="#item_ramp"><CODE>ramp()</CODE></A> preforms smooth ramp activation between 0 and 1 if $r is 1,
or between -1 and 1 if $r is 2. $r defaults to 1.
<P>You can get this into your namespace with the ':acts' export
tag as so:
</P>
<PRE>
use AI::NeuralNet::Mesh ':acts';</PRE>
<P>Note: when using a <A HREF="#item_ramp"><CODE>ramp()</CODE></A> activatior, train the
net at least TWICE on the data set, because the first
time the <A HREF="#item_ramp"><CODE>ramp()</CODE></A> function searches for the top value in
the inputs, and therefore, results could flucuate.
The second learning cycle guarantees more accuracy.</P>
<P>No code to show here, as it is almost exactly the same as range().</P>
<P></P>
<DT><STRONG><A NAME="item_and_gate">and_gate($threshold);</A></STRONG><BR>
<DD>
Self explanitory, pretty much. This turns the node into a basic AND gate.
$threshold is used to decide if an input is true or false (1 or 0). If
an input is below $threshold, it is false. $threshold defaults to 0.5.
<P>You can get this into your namespace with the ':acts' export
tag as so:
</P>
<PRE>
use AI::NeuralNet::Mesh ':acts';</PRE>
<P>Let's look at the code real quick, as it shows how to get at the indivudal
input connections:</P>
<PRE>
= line 1 = sub {
= line 2 = my $sum = shift;
= line 3 = my $self = shift;
= line 4 = my $threshold = 0.50;
= line 5 = for my $x (0..$self->{_inputs_size}-1) {
= line 6 = return 0.000001 if(!$self->{_inputs}->[$x]->{value}<$threshold)
= line 7 = }
= line 8 = return $sum/$self->{_inputs_size};
= line 9 = }</PRE>
<P>Line 2 and 3 pulls in our sum and self refrence. Line 5 opens a loop to go over
all the input lines into this node. Line 6 looks at each input line's value
and comparse it to the threshold. If the value of that line is below threshold, then
we return 0.000001 to signify a 0 value. (We don't return a 0 value so that the network
doen't get hung trying to multiply a 0 by a huge weight during training [it just will
keep getting a 0 as the product, and it will never learn]). Line 8 returns the mean
value of all the inputs if all inputs were above threshold.</P>
<P>Very simple, eh? :)
</P>
<P></P>
<DT><STRONG><A NAME="item_or_gate">or_gate($threshold);</A></STRONG><BR>
<DD>
<P>Self explanitory. Turns the node into a basic OR gate, $threshold is used same as above.</P>
<P>You can get this into your namespace with the ':acts' export
tag as so:
</P>
<PRE>
use AI::NeuralNet::Mesh ':acts';</PRE>
<P></P></DL>
<P>
<HR>
<H1><A NAME="variables">VARIABLES</A></H1>
<DL>
<DT><STRONG><A NAME="item_%24AI%3A%3ANeuralNet%3A%3AMesh%3A%3AConnector">$AI::NeuralNet::Mesh::Connector</A></STRONG><BR>
<DD>
This is an option is step up from average use of this module. This variable
should hold the fully qualified name of the function used to make the actual connections
between the nodes in the network. This contains '_c' by default, but if you use
this variable, be sure to add the fully qualified name of the method. For example,
in the ALN example, I use a connector in the main package called <CODE>tree()</CODE> instead of
the default connector. Before I call the <A HREF="#item_new"><CODE>new()</CODE></A> constructor, I use this line of code:
<PRE>
$AI::NeuralNet::Mesh::Connector = 'main::tree'
</PRE>
<P>The tree() function is called as a blessed method when it is used internally, providing
access to the bless refrence in the first argument. See notes on CUSTOM NETWORK CONNECTORS,
below, for more information on creating your own custom connector.</P>
<P></P>
<DT><STRONG><A NAME="item_%24AI%3A%3ANeuralNet%3A%3AMesh%3A%3ADEBUG">$AI::NeuralNet::Mesh::DEBUG</A></STRONG><BR>
<DD>
This variable controls the verbosity level. It will not hurt anything to set this
directly, yet most people find it easier to set it using the <A HREF="#item_debug"><CODE>debug()</CODE></A> method, or
any of its aliases.
<P></P></DL>
<P>
<HR>
<H1><A NAME="custom network connectors">CUSTOM NETWORK CONNECTORS</A></H1>
<P>Creating custom network connectors is step up from average use of this module.
However, it can be very useful in creating other styles of neural networks, other
than the default fully-connected feed-foward network.</P>
<P>You create a custom connector by setting the variable $AI::NeuralNet::Mesh::Connector
to the fully qualified name of the function used to make the actual connections
between the nodes in the network. This variable contains '_c' by default, but if you use
this variable, be sure to add the fully qualified name of the method. For example,
in the ALN example, I use a connector in the main package called <CODE>tree()</CODE> instead of
the default connector. Before I call the <A HREF="#item_new"><CODE>new()</CODE></A> constructor, I use this line of code:</P>
<PRE>
$AI::NeuralNet::Mesh::Connector = 'main::tree'
</PRE>
<P>The tree() function is called as a blessed method when it is used internally, providing
access to the bless refrence in the first argument.</P>
<P>Example connector:</P>
<PRE>
sub connect_three {
my $self = shift;
my $r1a = shift;
my $r1b = shift;
my $r2a = shift;
my $r2b = shift;
my $mesh = $self->{mesh};
for my $y (0..($r1b-$r1a)-1) {
$mesh->[$y+$r1a]->add_output_node($mesh->[$y+$r2a-1]) if($y>0);
$mesh->[$y+$r1a]->add_output_node($mesh->[$y+$r2a]) if($y<($r2b-$r2a));
$mesh->[$y+$r1a]->add_output_node($mesh->[$y+$r2a+1]) if($y<($r2b-$r2a));
}
}</PRE>
<P>This is a very simple example. It feeds the outputs of every node in the first layer
to the node directly above it, as well as the nodes on either side of the node directly
above it, checking for range sides, of course.</P>
<P>The network is stored internally as one long array of node objects. The goal here
is to connect one range of nodes in that array to another range of nodes. The calling
function has already calculated the indices into the array, and it passed it to you
as the four arguments after the $self refrence. The first two arguments we will call
$r1a and $r1b. These define the start and end indices of the first range, or ``layer.'' Likewise,
the next two arguemnts, $r2a and $r2b, define the start and end indices of the second
layer. We also grab a refrence to the mesh array so we dont have to type the $self
refrence over and over.</P>
<P>The loop that folows the arguments in the above example is very simple. It opens
a <CODE>for()</CODE> loop over the range of numbers, calculating the size instead of just going
$r1a..$r1b because we use the loop index with the next layer up as well.</P>
<P>$y + $r1a give the index into the mesh array of the current node to connect the output FROM.
We need to connect this nodes output lines to the next layers input nodes. We do this
with a simple method of the outputing node (the node at $y+$r1a), called add_output_node().</P>
<P><CODE>add_output_node()</CODE> takes one simple arguemnt: A blessed refrence to a node that it is supposed
to output its final value TO. We get this blessed refrence with more simple addition.</P>
<P>$y + $r2a gives us the node directly above the first node (supposedly...I'll get to the ``supposedly''
part in a minute.) By adding or subtracting from this number we get the neighbor nodes.
In the above example you can see we check the $y index to see that we havn't come close to
any of the edges of the range.</P>
<P>Using $y+$r2a we get the index of the node to pass to <CODE>add_output_node()</CODE> on the first node at
$y+<STRONG>$r1a</STRONG>.</P>
<P>And that's all there is to it!</P>
<P>For the fun of it, we'll take a quick look at the default connector.
Below is the actual default connector code, albeit a bit cleaned up, as well as
line numbers added.</P>
<PRE>
= line 1 = sub _c {
= line 2 = my $self = shift;
= line 3 = my $r1a = shift;
= line 4 = my $r1b = shift;
= line 5 = my $r2a = shift;
= line 6 = my $r2b = shift;
= line 7 = my $mesh = $self->{mesh};
= line 8 = for my $y ($r1a..$r1b-1) {
= line 9 = for my $z ($r2a..$r2b-1) {
= line 10 = $mesh->[$y]->add_output_node($mesh->[$z]);
= line 11 = }
= line 12 = }
= line 12 = }
</PRE>
<P>Its that easy! The simplest connector (well almost anyways). It just connects each
node in the first layer defined by ($r1a..$r1b) to every node in the second layer as
defined by ($r2a..$r2b).</P>
<P>Those of you that are still reading, if you do come up with any new connection functions,
PLEASE SEND THEM TO ME. I would love to see what others are doing, as well as get new
network ideas. I will probably include any connectors you send over in future releases (with
propoer credit and permission, of course).</P>
<P>Anyways, happy coding!</P>
<P>
<HR>
<H1><A NAME="what can it do">WHAT CAN IT DO?</A></H1>
<P>Rodin Porrata asked on the ai-neuralnet-backprop malining list,
``What can they [Neural Networks] do?''. In regards to that questioin,
consider the following:</P>
<P>Neural Nets are formed by simulated neurons connected together much the same
way the brain's neurons are, neural networks are able to associate and
generalize without rules. They have solved problems in pattern recognition,
robotics, speech processing, financial predicting and signal processing, to
name a few.</P>
<P>One of the first impressive neural networks was NetTalk, which read in ASCII
text and correctly pronounced the words (producing phonemes which drove a
speech chip), even those it had never seen before. Designed by John Hopkins
biophysicist Terry Sejnowski and Charles Rosenberg of Princeton in 1986,
this application made the Backprogagation training algorithm famous. Using
the same paradigm, a neural network has been trained to classify sonar
returns from an undersea mine and rock. This classifier, designed by
Sejnowski and R. Paul Gorman, performed better than a nearest-neighbor
classifier.</P>
<P>The kinds of problems best solved by neural networks are those that people
are good at such as association, evaluation and pattern recognition.
Problems that are difficult to compute and do not require perfect answers,
just very good answers, are also best done with neural networks. A quick,
very good response is often more desirable than a more accurate answer which
takes longer to compute. This is especially true in robotics or industrial
controller applications. Predictions of behavior and general analysis of
data are also affairs for neural networks. In the financial arena, consumer
loan analysis and financial forecasting make good applications. New network
designers are working on weather forecasts by neural networks (Myself
included). Currently, doctors are developing medical neural networks as an
aid in diagnosis. Attorneys and insurance companies are also working on
neural networks to help estimate the value of claims.</P>
<P>Neural networks are poor at precise calculations and serial processing. They
are also unable to predict or recognize anything that does not inherently
contain some sort of pattern. For example, they cannot predict the lottery,
since this is a random process. It is unlikely that a neural network could
be built which has the capacity to think as well as a person does for two
reasons. Neural networks are terrible at deduction, or logical thinking and
( run in 0.582 second using v1.01-cache-2.11-cpan-39bf76dae61 )