J.L. Blue, ACMD
The Image Recognition group of the Computer Systems Laboratory, has been actively developing and testing character recognition systems for the IRS and Census and fingerprint recognition systems for the FBI. I serve as a mathematical and algorithmical consultant to the group.
In classifying characters, for instance, one starts with a large corpus of characters with known classes. These characters may be used to determine the parameters in a neural network; the resulting network may be used for classifying unknown characters. This year I found that a new activation function (a sinusoid instead of a sigmoid) reduces the classification error. When used in a pruned network, the error vs. reject curve is substantially improved --- for a given desired classification error, only one-fourth to one-half as many unknowns need to be rejected as undecidable.
This work is supported by the Computer Systems Laboratory.