Machine-Learning Algorithms with Low Labeling Cost: Semisupervised Learning, Active Learning, and Transfer Learning

瀏覽統計 Email 通知 RSS Feed

簡歷

基本資料

Project title

Code/計畫編號

MOST107-2221-E019-044

Translated Name/計畫中文名

低標示訓練樣本成本之機器學習演算法: 半監督式、主動式、遷移式學習

Project Coordinator/計畫主持人

Chin-Chun Chang

Funding Organization/主管機關

National Science and Technology Council

Department/Unit

Department of Computer Science and Engineering

Website

https://www.grb.gov.tw/search/planDetail?id=12675241

Year

2018

Start date/計畫起

01-08-2018

Expected Completion/計畫迄

31-07-2019

Bugetid/研究經費

544千元

ResearchField/研究領域

資訊科學--軟體

蒐集或產生具有標籤的訓練樣本通常是既耗時且耗人力的工作。在需要數量龐大具類別標籤訓練資料的機器學習應用上，這個困難就更顯嚴峻。半監督式學習、主動學習、遷移學習在不同的應用情境下，可以用來緩解這項問題。針對這三種機器學習方式，本計畫擬以兩年時間開發三個演算法。其中，半監督式學習的研究項目將提升資料相依核函數(data-dependent kernel)資料規模延展性 (scalability)；主動學習將發展開發非監督式學習結果並適用於大規模機器學習問題的演算法；遷移學習將使用生成對抗網路(generative adversarial networks)於域調適問題(domain adaptation problem)。擬研究的三個演算法皆可用較低的樣本標示成本來開發機器學習系統，將有助於機器學習系統的開發。Labeling a large amount of training samples for large-scale machine learning tasks is expensive. In different application scenarios, semi-supervised learning, active learning, and transfer learning can be applied to reduce demands for labeled training samples. In this two-year project, we shall develop three novel algorithms on these three kinds of machine-learning paradigm. For semi-supervised learning, we shall enhance the scalability of the data-dependent kernel. Regarding active learning, we shall develop a batch-mode active learning algorithm which exploits the result of unsupervised learning for large-scale machine learning applications. For transfer learning, we shall develop a novel structure of generative adversarial networks for domain adaptation problems. The three algorithms will be useful for developing machine-learning applications which have lack of labeled training samples.

Keyword(s)

半監督式學習
資料相依核函數
主動學習
遷移學習
生成對抗網路
semi-supervised learning
data-dependent kernels
active learning
transfer learning
generative adversarial networks

DSpace CRIS

Machine-Learning Algorithms with Low Labeling Cost: Semisupervised Learning, Active Learning, and Transfer Learning

基本資料

Description