Home // INTELLI 2021, The Tenth International Conference on Intelligent Systems and Applications // View article


Controlling Individual and Collective Information for Generating Interpretable Models of Multi-Layered Neural Networks

Authors:
Ryotaro Kamimura

Keywords: individual, collective, information, selectivity, partial compression, interpretation, generalization

Abstract:
The present paper aims to control selective information to understand the main mechanism of information processing in multi-layered neural networks. We propose two types of selective information, namely, individual and collective selective information, or simply, individual and collective information. The individual information represents to what degree a neuron is connected specifically to another one, and it should be increased as much as possible. Then, we try to use this abundant information as impartially as possible, reducing the specificity of collective neurons and reducing collective information. By controlling the ratio of individual and collective information, we can realize a number of different types of states to be interpreted, leading to the interpretation of the inference mechanism. The method was applied to the bankruptcy data set. In the experiments, we successfully increased individual information and decreased collective information. By examining partially compressed weights, we could see how neural networks, by controlling the selective information, can process information content in multi-layered neural networks. This examination of information flow can lead us to understand the main inference mechanism of neural networks.

Pages: 27 to 35

Copyright: Copyright (c) IARIA, 2021

Publication date: July 18, 2021

Published in: conference

ISSN: 2308-4065

ISBN: 978-1-61208-882-2

Location: Nice, France

Dates: from July 18, 2021 to July 22, 2021