Share Email Print
cover

Proceedings Paper

Optimal robustness of supervised learning from a noniterative point of view
Author(s): Chia-Lun John Hu
Format Member Price Non-Member Price
PDF $14.40 $18.00

Paper Abstract

In most artificial neural network applications, (e.g. pattern recognition) if the dimension of the input vectors is much larger than the number of patterns to be recognized, generally, a one- layered, hard-limited perceptron is sufficient to do the recognition job. As long as the training input-output mapping set is numerically given, and as long as this given set satisfies a special linear-independency relation, the connection matrix to meet the supervised learning requirements can be solved by a noniterative, one-step, algebra method. The learning of this noniterative scheme is very fast (close to real-time learning) because the learning is one-step and noniterative. The recognition of the untrained patterns is very robust because a universal geometrical optimization process of selecting the solution can be applied to the learning process. This paper reports the theoretical foundation of this noniterative learning scheme and focuses the result at the optimal robustness analysis. A real-time character recognition scheme is then designed along this line. This character recognition scheme will be used (in a movie presentation) to demonstrate the experimental results of some theoretical parts reported in this paper.

Paper Details

Date Published: 18 August 1995
PDF: 6 pages
Proc. SPIE 2622, Optical Engineering Midwest '95, (18 August 1995); doi: 10.1117/12.216859
Show Author Affiliations
Chia-Lun John Hu, Southern Illinois Univ./Carbondale (United States)


Published in SPIE Proceedings Vol. 2622:
Optical Engineering Midwest '95
Rudolph P. Guzik, Editor(s)

© SPIE. Terms of Use
Back to Top