NET-QUANTITIES:  Quantities from log files relating to networks.

The quantities below relating to neural network models can be obtained
from log files (eg, for use in net-plt).  Note that the generic
quantities documented in quantities.doc and the quantities relating to
Markov chain methods documented in mc-quantities.doc will also be
available.

    xn   Array of n'th input values for training cases
    Xn   Array of n'th input values for test cases
    in   Same as xn, except n must not be omitted
    In   Same as Xn, except n must not be omitted

    on   Array of n'th output values for training cases
    On   Array of n'th output values for test cases

    yn   Array of n'th "guessed" values for training cases
    Yn   Array of n'th "guessed" values for test cases

    gn   Array of n'th randomly "guessed" values for training cases
    Gn   Array of n'th randomly "guessed" values for test cases

    z[n] Array of n'th target values for training cases
    Z[n] Array of n'th target values for test cases
    tn   Same as zn, except n must not be omitted
    Tn   Same as Zn, except n must not be omitted

    P    Log prior probability of the network parameters with current 
         hyperparameter values

    l    Array of minus log probabilities for training cases.  If not 
         specified to be an array, the average minus log probability over 
         training cases.

    L    Array of minus log probabilities for test cases.  If not 
         specified to be an array, the average minus log probability over 
         test cases.

    a[n] Array of absolute errors of n'th target for training cases, 
         or sum of absolute errors if 'n' not specified.  If not 
         specified to be an array, the average error over training 
         cases.

    A[n] Array of absolute errors of n'th target for test cases, or 
         sum of absolute errors if 'n' not specified.  If not specified 
         to be an array, the average error over test cases.

    b[n] Array of squared errors of n'th target for training cases, or 
         sum of squared errors if 'n' not specified.  If not specified 
         to be an array, the average error over training cases.

    B[n] Array of squared errors of n'th target for test cases, or sum 
         of squared errors if 'n' not specified.  If not specified to 
         be an array, the average error over test cases.

    c    Expected classification error on training set, assuming that 
         the model is right.
 
    C    Expected classification error on test set, assuming that the
         model is right.

    vn   Array of std. dev. over training set of units in hidden layer 
         n; average std. dev. if scalar.

    Vn   Array of std. dev. over test set of units in hidden layer n;
         average std. dev. if scalar.

    hn   Value of n'th top-level (common) hyperparameter if scalar, or
         array of associated unit-level hyperparameters.  The indexes
         used are as displayed with net-display.  It is not valid to
         use the scalar form for a group of adjustments.  Noise 
         standard deviations cannot be accessed in this way - use 'n'.

    wn   Array of values for n'th group of parameters for network.  
         The indexes used are as displayed with net-display.

    Wn   Square root of average squared magnitude for n'th group of 
         parameters if scalar, or array of square roots of average 
         squared magnitudes of parameters in group associated with 
         each source unit.  The indexes used are as displayed with 
         net-display.

    Mn   Magnitudes of last hidden layer to output weights to n'th 
         output, sorted in decreasing order of size

    n    For regression model: Common noise std. dev. if scalar, 
         or array of output-specific noise levels.  For class model:
         expected entropy on training cases, assuming the model is right.

    N    For regression model: Common noise variance if scalar, or 
         array of output-specific noise levels.  For class model:
         expected entropy on test cases, assuming model is right.

When a value for 'n' is needed but not specified, it defaults to zero.
This isn't allowed with 'i', 'I', 't', and 'T', however, since they
have other meanings when 'n' is omitted.

The 'o' values are the values of the units in the final layer, before
interpretation by the model.  The 'y' values are the same as the 'o'
values for regression models.  For binary and multi-class models, the
'y' values are the bit/class probabilities obtained from the 'o'
values.  The 'g' values are taken randomly from the predictive
distribution.  For multi-class models, the 'a', 'A', 'b', and 'B'
quantities are based on the difference between the vector of
probabilities for the classes and the vector of 0s and 1s with a
single 1 indicating the correct class.

Depending on whether any training/test cases are available, and if
they are, whether they include targets, some of the above quantities
may not be defined.  The 'c' and 'C' quantities are not defined for
regression models.  The 'o' quantities are allowed with survival
models only when the hazard is constant (independent of time).  For
multi-class models, the 'z' and 'Z' quantities with a modifier ('n')
refer to the representation of the class as a binary vector with the
true class indicated by a one; with no modifier, they refer to the
numeric value of the class.  (Since 'n' can't be omitted for 't' and
'T', these synonyms can't be used for 'z' and 'Z' with no modifier.)

For censored survival data, the target for the 'A', 'a', 'B', 'b', 'Z',
and 'z' options is taken to be the censoring time, which is not very 
meaningful. 

The quantities relating to test cases use the test set specified using
data-spec by default.  A different test set can be used by including
arguments of the form

    / test-inputs [ test-targets ]

as the last arguments of the command that accesses these quantities
(eg, net-plt).

            Copyright (c) 1995-2004 by Radford M. Neal