부분 활성 신경망 분류기의 설계

Abstract

본 논문에서는 무리의 위치에 적응적으로 동작하고 높은 분류율, 자기조직적 기능을 유지하면서 내부노드 중 분류에 기여할 가능성이 있는 것 들 만을 활성화하는 PA-AEDC(Partially Activated-Adaptive Euclidean Distance Classifier) 신경망을 설계하였다. 이 신경망은 주신경망과 부신경망으로 구성되어 있으며, 주신경망은 분류의 일을, 부 신경만은 주신경망 중 일부분을 선택하여 활성화해 주는 역할을 한다.

성능을 파악하기 위하여 분류 실험을 해본 결과, 분류율은 AMDC나 AEDC와 비슷하게 나왔으나, 분류 시연산에 참여하는 노드 수는 PA-AEDC의 경우가 대폭 감소되었으며, 따라서 효과적인 동작이 가능하였다.

In this paper, PA-AEDC(Partially Activated-adaptive Euclidean Distance Classifier) neural network is designed. This one can adaptively operate for the position of clusters and can self-organize the internal nodes and their connections. PA-AEDC consists of two networks:the primary net and the secondary net. The primary net mainly acts as a feature vector classifier, while the secondary net selects and activates the output and internal nodes of the primary net which are supposed to contribute in the classification.

As a result of experiments of classification, we got the expected classifying rates as high as that of AMDC of AEDC. But the number of nodes of PA-AEDC that participate in the arithmetic operation was much reduced.

In this paper, PA-AEDC(Partially Activated-adaptive Euclidean Distance Classifier) neural network is designed. This one can adaptively operate for the position of clusters and can self-organize the internal nodes and their connections. PA-AEDC consists of two networks:the primary net and the secondary net. The primary net mainly acts as a feature vector classifier, while the secondary net selects and activates the output and internal nodes of the primary net which are supposed to contribute in the classification.

As a result of experiments of classification, we got the expected classifying rates as high as that of AMDC of AEDC. But the number of nodes of PA-AEDC that participate in the arithmetic operation was much reduced.