生成式對抗網路架構搜尋;GANAS: Generative Adversarial Network Architecture Search

NCUIR > School of Management at National Central University > Graduate Institute of Information Management > Electronic Thesis & Dissertation > Item 987654321/84006

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/84006

Title:	生成式對抗網路架構搜尋;GANAS: Generative Adversarial Network Architecture Search
Authors:	吳佳臻;WU, CHIA-CHEN
Contributors:	資訊管理學系
Keywords:	機器學習;深度學習;神經架構搜索;生成式對抗網路;Machine Learning;Deep Learning;Neural Architecture Search;Generative Adversarial Networks
Date:	2020-07-15
Issue Date:	2020-09-02 17:54:19 (UTC+8)
Publisher:	國立中央大學
Abstract:	近年來，機器學習在各個領域得到了廣泛的應用，並取得了突出的成績。在機器學習領域的各種方法中，深度學習是最受關注的一種，深度學習能夠快速處理海量資訊，滲透並改變著我們的日常生活。在深度學習中，設計優秀的神經網路架構是非常重要的事情，然而，設計一個優秀的架構不僅需要深度學習以及相關領域的專業知識，還需要在目標任務領域有足夠的經驗。因此，目前有很多關於自動產生神經網路架構的研究並實現神經網路的自動設計，然而大多數的研究非常消耗計算資源。因此在本文中，我們提出了一種新的方法，並將其稱為生成對抗式網路架構搜索（GANAS），該模型將conditional GANs擴展到NAS領域，最終目標是使用訓練有素的生成器生成神經網路架構，這種方法最大的特點是根據不同的資料會產生不同的神經網路架構，從而省去人工手動設計神經網路架構的時間，同時，只需要少量的運算資源便可達成我們的任務。;In recent years, machine learning has gained a wide range of applications in various fields and has achieved outstanding results. Among the various approaches in the machine learning field, deep learning is the one that has received the most attention; deep learning can process vast amounts of information quickly, permeates, and changes our daily lives. In deep learning, designing excellent neural network architecture is very important, however, designing an excellent architecture requires not only deep learning and expertise in the relevant field but also sufficient experience in the target task area. Therefore, there is a lot of research on generating neural network architectures automatically, however, such search methods are very consuming computing resources. Therefore, in this paper, we propose a new approach and call GANAS, the model extends conditional GANs into the realm of NAS, with the ultimate goal of generating neural network architectures using trained well generators, the best feature of this method is that different neural network architectures are generated according to the data, thus saving the time of designing network architectures, and at the same time, only a small amount of computational resources is needed to achieve our task.
Appears in Collections:	[Graduate Institute of Information Management] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	79	View/Open

社群 sharing

Loading...