請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/6375
標題: | 知識轉移於智慧家庭環境之行為辨識應用 Building Activity Recognition Models Using Transfer Learning in Smart Home Environment |
作者: | Yi-ting Chiang 蔣益庭 |
指導教授: | 許永真 |
關鍵字: | 傑森-向農偏差,轉移學習,行為辨識,機器學習,智慧家庭, Jenson-Shannon divergence,transfer learning,activity recognition,machine learning,smart home, |
出版年 : | 2013 |
學位: | 博士 |
摘要: | 大多數的行為辨識研究應用了機器學習方法。傳統上,機器學習基於一個假設來建立模型:用來建立模型時的環境與應用此模型的環境是相同的。此假設並非總是成立。若是可以先於一實驗室環境收集資料,並利用轉移學習(Transfer learning)來減低收集資料的花費,建立行為辨識模型於一智慧家庭將更為實用。
轉移學習移除了建立及應用模型時環境必須相同的限制,允許學習及測試模型的資料庫可以相異,並成功地應用於許多機器學習問題。此研究介紹一應用於行為辨識的知識轉移架構。具體來說,我們定義一個以特徵為基礎,可以自動計算新的特徵表述來轉移知識的知識轉移架構。我們於下面兩種情境下實際建造行為辨識模型來驗證此架構。情境一假設資料已標記的來源領域資料庫以及關於來源及目標領域資料庫的背景知識皆可知,但目標資料庫並不可得;情境二假設資料已標記的來源以及目標領域資料庫皆可得。實驗證明此知識轉移架構可在兩相異環境中成功地萃取並轉移知識。 Most activity recognition research makes use of machine learning methods. Traditional machine methods build a model under the assumption that this model will be applied to the same environment where the training dataset is collected, which is sometimes unrealistic in real world. In addition, collecting data sets in every environment where the model is going to be applied is not feasible. Building activity recognition models in a smart home is more practical if we can collect the dataset in a laboratory environment, and use transfer learning to reduce the effort of data collection. Transfer learning relaxes this constraint, so that the training and the testing dataset can be different. It has considerable success in many machine learning problems. In this work, we propose a knowledge transfer framework for activity recognition. Specifically, we propose a feature-based knowledge transfer framework which can automatically find the new formulation of features to transfer knowledge between two domains. We apply this framework under two different scenarios to train activity recognition models: In the first scenario, a labelled source dataset is available, and we have the background knowledge about the source and target domain, but we do not have any target domain data samples. In the second scenario, we have both labelled the source and target domain dataset. The experimental results show that this framework can successfully extract and transfer knowledge between two different domains. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/6375 |
全文授權: | 同意授權(全球公開) |
顯示於系所單位: | 資訊工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-102-1.pdf | 1.6 MB | Adobe PDF | 檢視/開啟 |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。