
Unsupervìsed Training
337
acceptable is to watch the maximum (across all neurons) length of the
mean correction vector. (This quantity is discussed in the Convergence
section.) The maximum correction should quickly and steadily
decrease. If it only slowly decreases, the learning rate is too small. If
it occasionally (or regularly) increases, the learning rate is dangerously
high, leading to instability. Note that the above advice applies only if
the corrections are averaged across the entire training set, which is
generally recommended. If weights are corrected in response to each
training case, the correction vectors can be expected to var ...