Encodes a categorical values with an unbounded vocabulary. Values are encoding by incrementing a
few locations in the output vector with a weight that is either defaulted to 1 or that is looked
up in a weight dictionary. By default, only one probe is used which should be fine but could
cause a decrease in the speed of learning because more features will be non-zero. If a large
feature vector is used so that the probability of feature collisions is suitably small, then this
can be decreased to 1. If a very small feature vector is used, the number of probes should
probably be increased to 3.