Performance Enhancement of Deep Learning with Hidden Parameter Inductions by Grafting a Pre-Defined Computation Module

SEI Makoto, UTSUMI Akira, YAMAZOE Hirotake, LEE Joo-Ho

doi:10.14923/transinfj.2022pdp0026

Bibliographic Information

Other Title

事前定義計算モジュールの組込みによる隠れパラメータの誘導と深層学習の性能向上

Abstract

This paper proposes a method of integrating known relations about the target task into a neural network as predefined computational modules. During the training process, the network is automatically induced to estimate the hidden parameters required by the embedded computational modules. In experiments, the proposed method was applied to 2D eyeball center position estimation. Experimental results confirmed that the accuracy of 2D eyeball center position estimation is improved and that the system acquires the ability to estimate 3D eye positions, which are hidden parameters.

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 J106-D (4), 268-276, 2023-04-01

The Institute of Electronics, Information and Communication Engineers

Keywords

Details 詳細情報について

CRID: 1390858529773433984

DOI: 10.14923/transinfj.2022pdp0026

ISSN: 18810225; 18804535

Text Lang: ja

Data Source

JaLC
KAKEN

Abstract License Flag: Disallowed

Export