Home // COGNITIVE 2016, The Eighth International Conference on Advanced Cognitive Technologies and Applications // View article


Self-Organized Potential Competitive Learning to Improve Interpretation and Generalization in Neural Networks

Authors:
Ryotaro Kamimura
Ryozo Kitajima
Osamu Uchida

Keywords: Self-organizing maps; Potentiality; Interpretation; Generalization.

Abstract:
The present paper proposes a new learning method called “self-organized potential competitive learning” to improve generalization and interpretation performance. In this method, the self-organizing map (SOM) is used to produce knowledge (SOM knowledge) on input patterns. By considering the potentiality of neurons rather than stored information, it can be used to train supervised learning. Highly potential neurons are supposed to respond to as many input patterns and neurons as possible. This property is, for the first approximation, described by the variance of connection weights. The method was applied to real second language learning data (Japanese learners of English) and showed improved generalization performance. In addition, two important input neurons with high potentiality were detected, both of which represented inanimate subjects. This implies that Japanese students have difficulty dealing with inanimate subjects when learning English as a second language. This finding corresponds with the established knowledge on second language learning. The present results affirm the possibility of SOM knowledge to be applied to many different situations.

Pages: 32 to 37

Copyright: Copyright (c) IARIA, 2016

Publication date: March 20, 2016

Published in: conference

ISSN: 2308-4197

ISBN: 978-1-61208-462-6

Location: Rome, Italy

Dates: from March 20, 2016 to March 24, 2016