HPAVC/HPAVC CD-ROM.iso/NEUCLS3.ZIP/NURN.ZP/DATAMN.HLP

Text File | 1993-01-04 | 6KB | 124 lines

Outline

1. Files Needed or Produced by Software
2. Training and Testing Data
 a. IANS Format
 b. Data Files Included With This Package
3. File and Neural Net Limitations



1. Files Needed or Produced by Software
 a. MLP and functional link neural networks typically have three types
    of files associated with them. These three types are:
  (1) The network structure file. For the MLP, this file specifies the
      number of network layers, the number of artificial neurons
      (called units) in each layer, and the number of the first layer
      which the third and fourth (if there is one) layers connect to.
      For the functional link net, this file contains the network degree P
      (usually an integer between 1 and 5), the number of network inputs N
      and the number of outputs, and the dimension of the multinomial vector,
      which is L = (N+P)!/(N!P!).
  (2) The weight file, which gives the gains or coefficients along
      paths connecting the various units.
  (3) The training or testing data file, which gives example inputs
      and outputs for network learning, or for testing after learning.

 b. The network structure files have the extension "top". You can create
    your own network structure files within the backpropagation, fast
    training and functional link programs, if you want. Consider the
    MLP structure file, Grng.top shown below.

           4
          16          20          10           4
           1           1           1

    It has 4 layers. The first layer has 16 inputs, which means that
    each training or testing pattern has 16 numbers. It has 20 units in
    the first hidden layer, where "hidden" means that it is not an input
    or output layer. It has 10 units in the second hidden layer.
    The output layer has 4 units, corresponding to the 4 possible
    decisions that the network can make about the 16  input numbers.
    The last line of "1s" means that layers 2, 3, and 4 connect up with
    layer 1, layers 1 and 2, and layers 1, 2, and 3 respectively. This
    network is "fully connected", meaning that each layer connects with
    all previous layers. Fully connected networks are more powerful than
    and train faster than non fully connected networks. The fully
    connected networks are almost always smaller than non fully connected
    networks which perform the same operation.

2. Training and Testing Data
 a. IANS Format
    All data files must be put into formatted, IANS form, which means
    that each pattern or feature vector is followed by the correct
    class number (class id). The data analysis and pre-processing
    option (number 2) puts raw data into the IANS format.

    You can type out the data files to examine them, and you can use
    these files with other neural net software. For example, consider
    the training data file, Xor, shown below.

            0.  0.  1
            0.  1.  2
            1.  0.  2
            1.  1.  1

    There are four training patterns with two inputs each. Patterns 1 and 4
    belong to class 1, as indicated by the class number, 1, at the end of
    the first and last rows. The middle two rows or training patterns
    belong to class 2.

    The software can convert this file into either of two forms during
    training or testing. If the network is to have coded output, the number
    of output units is about log to the base 2 of the number of classes. For
    the file Xor for example, log2(2) = 1, so topology file Xor.top has 1
    output unit. If we were to use file Xor to train a network having uncoded
    outputs, then the number of outputs would equal the number of classes.
    Since Xor, Par, and Par4 are used in coded output networks in our demos,
    the corresponding networks have one output each. Since file Grng has
    four classes and is used in uncoded output networks, the networks for
    it have four outputs.

    Internal to the software, file XOR is converted to the form

            0.  0.  0
            0.  1.  1
            1.  0.  1
            1.  1.  0

     if the coded format is specified, where the third row stores the
     desired outputs for the corresponding patterns. If uncoded format
     is used, the file would be converted to the form

             0.  0.  0  1
             0.  1.  1  0
             1.  0.  1  0
             1.  1.  0  1

     For pattern number 1 for example, the first output unit has the desired
     value of 0, corresponding to class 1.

 b. Data Files Included With This Package

    The XOR data file, which corresponds to exclusive or, has 4 patterns,
       2 classes, and 2 inputs.
    The PAR4 data file, which corresponds to 4-input parity check, has 16
       patterns, 2 classes, and 4 inputs.
    The GRNG data file, which corresponds to recognition of 4 geometric
       shapes, has 4 classes, 800 vectors or patterns, and 16 inputs.
    The Gongtrn data file, which corresponds to recognition of handprinted
       numerals, has 10 classes, 3,000 vectors or patterns, and 16 inputs
       or features calculated from 32 by 24 pixel binary images.


3. File and Neural Net Limitations
 There is no limitation on data file size.
 MLP neural nets are limited to 40 or fewer units in each layer,
   including the input layer, one or two hidden layers, and the output
   layer.
 Functional link networks are limited to 40 inputs, 15 outputs, and
   5th degree.
 Conventional clustering and self-organizing map clustering are limited
   to 32 elements per vector and 2,048 clusters. There is no limit on
   the number of input patterns.