The simplest version of the RemoveUseless operation was
described in Section 2.7 
--any edges or states
whose counts are zero after training are removed.
Models can be pruned more aggressively by removing any states or edges
whose counts are less than some constant.
Pruning off edges that have only been used once results in fairly mild
pruning that usually does not change the average cost for the seed
sequences much, especially if retraining is done to tune the new path
through the model now used by the sequence that previously used the
deleted edge. More aggressive pruning can make much larger changes in
costs.