faq.html

All the tool for build a network able to reconize any shapes. Very complete, a good base for anythi

                    previously saved weights, saving current weights, feeding 
                    forward, calculating errors and for all kind of training 
                    methods. It has a sub class called 'Layer'. This simply 
                    represents a layer and facilitates calculations during feed 
                    forward and back propagation.<br>
                    <br>
                    <a href="#top">top</a><br>
                    <br>
                    <u><b><a name="16"></a>What is the function of the class 'Pattern'?</b></u><br>
                    Represents a single training pattern including input data 
                    and target data. If you use PatternSet class, you don't have 
                    to deal with this class. A PatternSet class includes one or 
                    more pattern objects.<br>
                    <br>
                    <a href="#top">top</a><br>
                    <br>
                    <u><b><a name="17"></a>What is the function of the class 'PatternSet'?</b></u><br>
                    Represents a set of patterns to be used during training. You 
                    can use all patterns for training or you can spare some part 
                    of it for cross validation and testing.<br>
                    <br>
                    <a href="#top">top</a><br>
                    <br>
                    <u><b><a name="18"></a>What is the function of the class 'Randomizer'?</b></u><br>
                    It generally coordinates random number generation. You can 
                    create an instance using a seed ( the code will produce the 
                    same sequence of random values each time you run the code ) 
                    or without seed (the code will use system clock, the 
                    sequence will not be the same).<br>
                    <br>
                    <a href="#top">top</a><br>
                    <br>
                    <u><b><a name="19"></a>What is the function of the class 'LineReader'?</b></u><br>
                    This class facilitates the file reading process. We use it 
                    for reading pattern files, configuration files, weight 
                    files.<br>
                    <br>
                    <a href="#top">top</a><br>
                    <br>
                    <u><b><a name="20"></a>What is 'cross validation data'?</b></u><br>
                    Neural networks can be over trained to the point where 
                    performance on new data actually deteriorates. Roughly 
                    speaking, overtraining results in a network that memorizes 
                    the individual exemplars, rather than trends in the data set 
                    as a whole. Cross validation is a process whereby part of 
                    the data set is set aside for the purpose of monitoring the 
                    training process, to guard against overtraining. (*)<br>
                    <br>
                    <a href="#top">top</a><br>
                    <br>
                    <u><b><a name="21"></a>What is 'test data'?</b></u><br>
                    The testing set is used to test the performance of the 
                    network. Once the network has been trained, the weights are 
                    then frozen, the testing set is fed into the network, and 
                    the network output is compared with the desired output. (*)<br>
                    <br>
                    <a href="#top">top</a><br>
                    <br>
                    <u><b><a name="22"></a>What is the function of the methods 'CrossValErrorRatio' 
                    and 'TestValErrorRatio'?</b></u><br>
                    Using these methods you can get a linear figure of what the 
                    error level is, based on cross validation data or test data. 
                    Since it is divided by average deviation, it is independent 
                    of the scale of the outputs.<br>
                    <br>
                    <a href="#top">top</a><br>
                    <br>
                    <u><b><a name="23"></a>Can I feed my own cross validation and test patterns?</b></u><br>
                    Yes you can. If you don't want patterns to be selected 
                    randomly and wish to create your own set of cross validation 
                    patterns out of your own file, you have to create a 
                    PatternSet object using '1' as ratiocrossval and '0' for 
                    other two ratios. Then you can use 
                    CrossValErrorRatio(yourpatternset) method to calculate the 
                    error. The same is true for test data (see example 2).<br>
                    <br>
                    <a href="#top">top</a><br>
                    <br>
                    <u><b><a name="24"></a>How can I set random weights after I create a net?</b></u><br>
                    After you use any of the constructors and create a net, all 
                    weights will be random values uniformly distributed between 
                    -1 and 1. They will remain so unless you train the network 
                    or load weights from a file.<br>
                    <br>
                    <a href="#top">top</a><br>
                    <br>
                    <u><b><a name="25"></a>What is a configuration file and what is the format of 
                    a configuration file?</b></u><br>
                    A configuration file is a saved configuration of a net. It 
                    is a text file, so it can be edited manually using a basic 
                    text editor. It includes the topology of the net as well as 
                    detailed information about units. It doesn't include 
                    information about weights though.<br>
                    Below is a configuration file taken from example 1.<br>
                    <br>
                    // First of all, we determine total number of units in the 
                    net.<br>
                    #neurons;5<br>
                    // Then we describe input units. There will be three fields:<br>
                    // type (for input units this will be 'i'); ID (a unique 
                    number for each unit);<br>
                    // layer (inputs are all in layer 0)<br>
                    i;0;0<br>
                    i;1;0<br>
                    // Now we describe hidden layers.<br>
                    // type;ID;layer;flatness;axon family;momentum;learning rate<br>
                    // type,ID,layer (all of those have the same meaning as in 
                    input layer)<br>
                    // type should be 'n' for hidden and output layers.<br>
                    // flatness: how flat is the sigmoid curve? The bigger it 
                    is, the flatter is the curve.<br>
                    // axon family: 't' tanh, 'g' logistic, 'l' linear.<br>
                    // momentum: momentum rate.<br>
                    // learning rate: (see 'Can I use different learning rates 
                    for each layer?')<br>
                    n;2;1;1;t;0.5;1<br>
                    n;3;1;1;t;0.5;1<br>
                    // output layer (same format as hidden layers)<br>
                    n;4;3;1;l;0.5;1<br>
                    // And now we describe synapses. For each connection between 
                    units,<br>
                    // there will be a line here.<br>
                    // First determine the number of synapses.<br>
                    #synapses;6<br>
                    // type; ID; source unit; target unit<br>
                    // type will be 's' for all synapses.<br>
                    // ID is a unique number for each synapse)<br>
                    // source unit is the ID of the source unit of this synapse.<br>
                    // target unit is the ID of the target unit of this synapse.<br>
                    s;0;0;2<br>
                    s;1;0;3<br>
                    s;2;1;3<br>
                    s;3;2;4<br>
                    s;4;3;4<br>
                    s;5;1;4<br>
                    <br>
                    <a href="#top">top</a><br>
                    <br>
                    <u><b><a name="26"></a>What is a pattern file and what is the format of a 
                    pattern file?</b></u><br>
                    A pattern file is a text file and it contains training 
                    patterns. If you use the constructor in PatternSet class, it 
                    will read values from a pattern file (examples 1 and 2). The 
                    values in a pattern file should be separated by semicolons 
                    ';'. This is the format of a 'csv' file. You can easily 
                    create one using MS Excel for example. A line in a pattern 
                    file includes input value(s) and target value(s).<br>
                    The line below was taken from 'example1.csv'. Here, first 
                    two values are input values, the last one is a target value.<br>
                    -0.787338;-0.028483;-0.815821<br>
                    <br>
                    <a href="#top">top</a><br>
                    <br>
                    <u><b><a name="27"></a>What is a weight file and what is the format of a 
                    weight file?</b></u><br>
                    The weights of a net are stored in a weight file. This is a 
                    text file. 'LoadWeights' and 'SaveWeights' methods in the 
                    class 'NeuralNet' loads and saves network weights (see 
                    examples). A weight file contains semicolon ';' separated 
                    values. Each line has three values: Type, ID, weight value.<br>
                    'Type' determines whether this is a synapse weight or a 
                    threshold. It is 'w' for the first and 't' for the second.<br>
                    'ID' determines the ID of the synapse in question (if this 
                    is a synapse weight) or the ID of the unit (if this is a 
                    threshold)<br>
                    'weight value' is the weight itself.<br>
                    Below is an example of this.<br>
                    w; 0; 1.9678931117918397<br>
                    w; 1; 0.13815102250339562<br>
                    w; 2; -2.0168124457485814<br>
                    t; 7; 0.4776425534517153<br>
                    t; 8; 1.4714782668522086<br>
                    t; 9; 0.13046564839824953<br>
                    <br>
                    <a href="#top">top</a><br>
                    <br>
                    <b><u><a name="28"></a>Do you give the source code for free?</u></b><br>
                    Yes. But please contact me if you wish to use it entirely or 
                    partially in any kind of project so that I can reference it. 
                    Please don't delete the top lines of the codes so that other 
                    people can reach that information too.<br>
                    <br>
                    <a href="#top">top</a><br>
                    <br>
                    <u><b><a name="29"></a>How can I get more detailed documentation?</b></u><br>
                    You can only find it with the comments in the code or you 
                    can ask me.<br>
                    <br>
                    <a href="#top">top</a><br>
                    <br>
                    <u><b><a name="30"></a>In which development environment has this package been 
                    tested?</b></u><br>
                    It is tested in Sun Java 2 SDK 1.4.0<br>
                    <br>
                    <a href="#top">top</a><br>
                    <br>
                    <u><b><a name="31"></a>How can I compile the code?</b></u><br>
                    I guess that you can compile using any java compiler. I 
                    personally use command line interface using Windows 2000 and 
                    Sun Java SDK 1.4.0. All you have to do is to uncompress all 
                    the files in the same directory and type for example:<br>
                    javac example1.java <br>
                    in the command line. Note that this is the way it works in 
                    Microsoft based operating systems and that I don't have 
                    experience on others. However, I don't think that there will 
                    be much difference, since we are talking about command line 
                    interfaces.<br>
                    <br>
                    <a href="#top">top</a><br>
                    <br>
                    <u><b><a name="32"></a>Error management in this package is not coded 
                    properly. Are you aware of it?</b></u><br>
                    Yes I am. I am planning to revise them in the future. Until 
                    then please be careful when you determine parameters.<br>
                    <br>
                    <a href="#top">top</a><br>
                    <br>
                    <u><b><a name="33"></a>Where can I find basic information about neural 
                    networks?</b></u><br>
                    You can try:<br>
                    http://www.shef.ac.uk/psychology/gurney/notes/contents.html<br>
                    ftp://ftp.sas.com/pub/neural/index.html<br>
                    http://directory.google.com/Top/Computers/Artificial_Intelligence/Neural_Networks/<br>
                    <br>
                    <a href="#top">top</a><br>
                    <br>
                    <u><b><a name="34"></a>How can I learn Java?</b></u><br>
                    I would recommend http://java.sun.com. It contains a lot of 
                    information and an excellent java tutorial.<br>
                    <br>
                    <a href="#top">top</a><br>
                    <br>
                    (*) Taken from NeuroSolutions 4.20 (NeuroDimension Inc.) 
                    Help document.</font></td>
			</tr>
			<tr>
				<td align="left">
					<a href="index.html">Home</a></td>
			</tr>
			</table>
	      </center>
        </div>
	</body>
</html>