12.8.1 Summed-Error Statistics

The SE_Stat object is a stat that iterates over the units that have target values in a network, and computes the difference between the activation and the target value. This is useful for monitoring learning performance over training. The current value of this statistic can be found in the se member. Also, there is a tolerance parameter which causes absolute differences of less than this amount to result in zero error. Thus, if one only was interested in whether the network was on the right side of .5, you would set the tolerance to .5 (assuming a 0 to 1 activation range).

There is also a CE_Stat and a RBpSE_Stat defined in the Bp version of the executable. These compute the cross-entropy error statistic and a version of squared-error that takes into account the dt parameter of the recurrent backprop algorithm.