The Mule Problem

Another problem is how can the network learn a new concept that is close to an old one.
For example, the network has been presented a range of horses and has thus has learned the horse concept.
The system is presented with a Mule.
As the Mule is really similar to a Horse (and there is no Mule concept) the Horse is activated.
This is good because the system can treat a Mule largely as a Horse.
However, if the system now is presented with a lot of Mules, it should be able to learn the concept Mule.
Unfortunately, each time a Mule is presented, Horse is activated, so the system merely enlarges its concept of Horse to include the extra features/neurons of the Mule.