By combining Kohonen learning and Grossberg learning a new type of mapping neural network is obtained. This counterpropagation network (CPN) functions as a statistically optimal self-programming lookup table. The paper begins with some introductory comments, followed by the definition of the CPN. Then a closed-form formula for the error of the network is developed. The paper concludes with a discussion of CPN variants and comments about CPN convergence and performance. References and a neurocomputing bibliography with a combined total of eighty entries are provided.
© 1987 Optical Society of America
Original Manuscript: July 18, 1987
Published: December 1, 1987
Robert Hecht-Nielsen, "Counterpropagation networks," Appl. Opt. 26, 4979-4984 (1987)