The TARGET architecture: a feature-oriented approach to connectionist word spotting

A new connectionist architecture with absolute classification capability is proposed. In the TARGET architecture, each unit has a target vector associated with it, which is the set of output values of units in a lower layer of the network which will cause the unit to be fully activated. When the outputs of all of the sending units closely match a unit´s target vector, the unit outputs a value close to zero. The network is trained by gradient descent, using a procedure derived in the same manner as the standard back propagation procedure. A rudimentary test of this system on the exclusive-or-problem is reported, in which a system achieves outputs accurate within 1%. A more extensive test of the system is reported, using a single-speaker isolated-word database of spelled Spanish words, with a vocabulary consisting of the 29 letters of the Spanish alphabet. The recognition rate using the new architecture was 94.0%, compared with 92.5% for standard backpropagation