Model Compression Using Multi-Trimmed Network Structure for Image Classification on Embedded Systems

Sarakon,  Pornthep

WEKO3

lat lon distance

[[sub_check.contents]]

[[sub_radio.contents]]

Field does not validate

[[sub_attr.contents]]　

インデックスツリー

アイテム

組み込みシステムにおける画像分類のためのマルチトリムネットワーク構造を用いたモデル圧縮

https://doi.org/10.18997/00008631

名前 / ファイル	ライセンス	アクション
kou_k_532.pdf (22.1 MB)

アイテムタイプ

学位論文 = Thesis or Dissertation(1)

公開日

2021-12-06

資源タイプ

資源タイプ識別子

http://purl.org/coar/resource_type/c_db06

資源タイプ

doctoral thesis

タイトル

Model Compression Using Multi-Trimmed Network Structure for Image Classification on Embedded Systems

言語

タイトル

組み込みシステムにおける画像分類のためのマルチトリムネットワーク構造を用いたモデル圧縮

言語

eng

著者

Sarakon, Pornthep

抄録

内容記述タイプ

Abstract

内容記述

Much effort has gone into developing smart robots, wherein perception and manipulation are among the most fundamental and challenging problems. Embedded systems （ESs） are critical in robot composition. However, as an embedded system, a robot brain has a fixed resource budget and is unsuitable for modern convolutional neural networks （CNNs）. Thus, the approach of CNN compression plays an important role in reducing their computational cost to make a suitable model for embedded systems. Recently, CNN compression approaches can be categorized into two groups, namely hand-crafted and model compression （MC） approach. The hand-crafted approach involves factorization and manual compression, but it is time consuming and usually requires significant amounts of manual effort and domain knowledge. Instead, the MC approach takes advantage of pre-trained models and it can solve a hand-crafted problem. The MC squeezes an existing model into one that is smaller and requires less computation. Although most MC methods can achieve a low latency or high accuracy, they are non-optimum accuracy–latency trade-off, complex, and do not affect certain dimensions （e.g., the width, resolution, and depth） of the models. To overcome this problem, the thesis presents a simple model-compression approach that optimize the accuracy–latency trade-off of the model. The multi-trimmed network structure （MTNS） is a robust combination of model compression （MC） techniques providing a lightweight model with trade-off optimization. The thesis describes a number of significant advances. Firstly, a new simple and efficient MC technique is introduced, which takes into width, resolution and depth compression. Secondly, a new multi-objective function is devised, which uses the accuracy–latency trade-off of compressed models to optimize the performance of a target model. Thirdly, a new training-accelerator is developed, which integrates pruning of convolutional kernels into shrinking the model structure to reduce training time at compressing width dimension. Finally, a new search strategy is developed, which combines Neural Architecture Search （NAS） with shrinking the model structure to explore more-complex conditions of shrinking the model structure with a relatively short training period. In an experimental evaluation, the thesis compares the performances of the proposed MTNS approach with those of CNN filter pruning, the model quantization technique, an adaptive mixture of low-rank factorizations, and knowledge distillation. The MTNS better resolved the accuracy–latency trade-off in image classification than the modern MC methods. It will be useful and friendly to the embedded system to perform a compressed model of MTNS with the maximum trade-off, lightweight, low computation and rapid process. The outstanding of the thesis is that the model compression problems have been solved by using MTNS techniques which are simple and optimum accuracy–latency trade-off for model compression.

内容記述タイプ

TableOfContents

内容記述

1 Introduction||2 Literature Reviews||3 Preliminary Knowledge and Technique for Model Compression||4 Shrinking Structure of Models||5 Shrinking Structure of Models with Training Accelerator||6 Trim Neural Architecture Search||7 Conclusions

備考

内容記述タイプ

Other

内容記述

九州工業大学博士学位論文学位記番号：工博甲第532号学位授与年月日：令和3年9月24日

キーワード

主題Scheme

Other

主題

Model compression

キーワード

主題Scheme

Other

主題

Computational efficiency

キーワード

主題Scheme

Other

主題

Image classification

キーワード

主題Scheme

Other

主題

Convolutional neural network

キーワード

主題Scheme

Other

主題

Embedded system

アドバイザー

河野, 英昭

学位授与番号

甲第532号

学位名

博士（工学）

学位授与年月日

2021-09-24

学位授与機関

学位授与機関識別子Scheme

kakenhi

学位授与機関識別子

17104

学位授与機関名

九州工業大学

学位授与年度

内容記述タイプ

Other

内容記述

令和3年度

出版タイプ

VoR

出版タイプResource

http://purl.org/coar/version/c_970fb48d4fbd8a85

アクセス権

open access

アクセス権URI

http://purl.org/coar/access_right/c_abf2

ID登録

10.18997/00008631

ID登録タイプ

JaLC

戻る

views

See details

	Views

Versions

Ver.1

2023-05-15 12:56:50.556547

Show All versions

Cite as

Other

エクスポート

OAI-PMH

JPCOAR 2.0
JPCOAR 1.0
DublinCore
DDI

Other Formats

インデックスリンク

インデックスツリー

アイテム

組み込みシステムにおける画像分類のためのマルチトリムネットワーク構造を用いたモデル圧縮

× Sarakon, Pornthep

Versions

Share

Cite as

Other

エクスポート

コミュニティ

メニューを最小化

インデックスリンク

インデックスツリー

アイテム

組み込みシステムにおける画像分類のためのマルチトリムネットワーク構造を用いたモデル圧縮

× Sarakon, Pornthep

Versions

Share

Cite as

Other

エクスポート

コミュニティ