Analysis of Feature Effectiveness Using a Selective Desensitization Neural Network
-
- SOMENO Shoichi
- University of Tsukuba
-
- HORIE Kazumasa
- University of Tsukuba
-
- MORITA Masahiko
- University of Tsukuba
Bibliographic Information
- Other Title
-
- 選択的不感化ニューラルネットによる特徴量の有効性の分析
Abstract
In function approximation, input features irrelevant to the output are known to lower the accuracy, but they are difficult to identify. Here we show that the selective desensitization neural network is highly robust to irrelevant features and that its connection weights have lower variance if the connections convey information only from irrelevant features. Numerical experiments showed that relevant features were distinguished based on the variance of connection weights even if they were uncorrelated with the output or relevant only in specific cases. These findings may lead to the development of new effective methods of feature selection and nonlinear data analysis.
Journal
-
- 電子情報通信学会論文誌D 情報・システム
-
電子情報通信学会論文誌D 情報・システム J102-D (8), 567-574, 2019-08-01
The Institute of Electronics, Information and Communication Engineers
- Tweet
Keywords
Details 詳細情報について
-
- CRID
- 1390282763131416448
-
- ISSN
- 18810225
- 18804535
-
- Text Lang
- ja
-
- Data Source
-
- JaLC
-
- Abstract License Flag
- Disallowed