Complete gradient optimization of a recurrent network applied to /b/,/d/,/g/ discrimination

A complete gradient optimization technique for connectionist networks with recurrent links is presented. A truncated gradient technique is available in the literature. A network with recurrent links for discrimination of /b/, /d/, and /g/ in the context of following /i/, /a/, or /u/ is designed. The performance of this network as optimized with either of the two techniques is then compared. It will be shown that the complete gradient is exact and simpler, and it will be demonstrated that it leads to superior performance.