Home // COGNITIVE 2023, The Fifteenth International Conference on Advanced Cognitive Technologies and Applications // View article
Authors:
Ryotaro Kamimura
Keywords: prototype network; comprehensive compression; interpretable compression; layer compression; collective compression; stabilizing compression; total de-compression; selective compression
Abstract:
The present paper aims to propose a method for creating an interpretable prototype network that is hidden within original multi-layered neural networks. The interpretation of the inference mechanism of neural networks has received much attention in recent times, leading to the development of various methods. However, these methods have focused on interpreting specific internal representations created by neural networks. There is an urgent need to propose an interpretation method that considers not only specific representations but also all internal representations created by a neural network, aiming for a more unified understanding of the fundamental properties of the inference mechanism. To address this problem, we propose the introduction of a prototype network that is hidden within the original multi-layered neural network. This is achieved through an interpretation method called ``comprehensive compression," which aims to replace the process of interpretation for finding a simple and interpretable prototype network. The method was applied to the analysis of customer data sets. The experimental results demonstrate that interpretable compression can simplify multi-layered neural networks and unify all obtained representations. It enables the detection and interpretation of non-linear as well as corresponding linear relations. The proposed method of the prototype network makes it possible to interpret not only specific instances but also a number of different instances. It helps us uncover the fundamental inference mechanism that is deeply hidden within neural networks.
Pages: 55 to 63
Copyright: Copyright (c) IARIA, 2023
Publication date: June 26, 2023
Published in: conference
ISSN: 2308-4197
ISBN: 978-1-68558-046-9
Location: Nice, France
Dates: from June 26, 2023 to June 30, 2023