Home // COGNITIVE 2025, The Seventeenth International Conference on Advanced Cognitive Technologies and Applications // View article

Neutralized Synchronic and Diachronic Potentiality for Interpreting Multi-Layered Neural Networks

Authors:
Ryotaro Kamimura

Keywords: diachronic; synchronic; di-synchronic; interpretation; potentiality; prototype

Abstract:
The present paper aims to propose a new method, called ``di-synchronic potentiality,'' to unify or neutralize diachronic and synchronic potentialities in order to seek for the simplest form of a network, referred to as the prototype. This method is necessary because the prototype is deeply hidden within surface networks, making it challenging to detect. The synchronic potentiality is measured by the complexity of connection weights at a specific learning time, while the diachronic potentiality is time-dependent. These potentialities tend to decrease as a natural property of learning. While this reduction is effective in eliminating unnecessary weights, it may also lead to the eventual elimination of important weights. The di-synchronic potentiality aims to unify or neutralize the reduction forces of diachronic and synchronic potentialities to increase synchronic potentiality. With this method, the synchronic potentiality does not necessarily increase, but at the very least, the reductive force is weakened or neutralized for solving the collision between the two types of reduction forces. The method was applied to artificial data simulating the bankruptcy of companies, with both linear and non-linear relations between inputs and targets. The results confirmed that there were strong and repeated forces striving to obtain the simplest form of potentiality. Ultimately, the method successfully produced representations with improved generalization, while simultaneously achieving the simplest relations between inputs and outputs for easier interpretation. From these results, we can conclude that detecting the simplest prototype can eventually lead to discovering more complex yet interpretable relations between inputs and outputs.

Pages: 17 to 25

Copyright: Copyright (c) IARIA, 2025

Publication date: April 6, 2025

Published in: conference

ISSN: 2308-4197

ISBN: 978-1-68558-260-9

Location: Valencia, Spain

Dates: from April 6, 2025 to April 10, 2025