Back error propagation - A critique

A neural-net algorithm known as back error propagation (BEP) is described and critiqued. A generic black-box view is presented, followed by a description of several noteworthy experiments. A review of the learning algorithm is given with a critique. Some training tips are offered along with a few variations on the basic design. A possible trend in neural net evolution is discussed, suggesting that the BEP could be heading toward a genetic dead end, despite early successes in small-scale experiments.<>