상세정보
Export to Refworks부가기능
Unsupervised Learning with Regularized Autoencoders
상세 프로파일
| 자료유형 | 학위논문 |
|---|---|
| 서명/저자사항 | Unsupervised Learning with Regularized Autoencoders. |
| 개인저자 | Zhao, Junbo. |
| 단체저자명 | New York University. Computer Science. |
| 발행사항 | [S.l.]: New York University., 2019. |
| 발행사항 | Ann Arbor: ProQuest Dissertations & Theses, 2019. |
| 형태사항 | 145 p. |
| 기본자료 저록 | Dissertations Abstracts International 81-05B. Dissertation Abstract International |
| ISBN | 9781392620069 |
| 학위논문주기 | Thesis (Ph.D.)--New York University, 2019. |
| 일반주기 |
Source: Dissertations Abstracts International, Volume: 81-05, Section: B.
Advisor: LeCun, Yann. |
| 이용제한사항 | This item must not be sold to any third party vendors. |
| 요약 | Deep learning has enjoyed remarkable successes in a variety of domains. These successes often emerge at the cost of large annotated datasets and training computationally heavy neural network models. The learning paradigm for this is called supervised learning. However, to reduce the sample complexity while improving the universality of the trained models is a crucial next step that may to artificial intelligence. Unsupervised Learning, in contrast to supervised learning, aims to build neural network models with more generic loss objectives requiring little or no labelling effort, and therefore it does not reside in any specific domain-task. In spite of the brevity of its goal, unsupervised learning is a broad topic that relates or includes several sub-fields, such as density estimation, generative modeling, world model, etc. In this thesis, we primarily adopt an energy-based view unifying these different fields (LeCun, 2006). The desired energy function reflects the data manifold by differentiating the energy assigned to the points on the data manifold against points off the manifold. Basing on this foundation, we first cast the popular autoencoder and adversarial learning framework into an energy-based perspective. Then, we propose several techniques or architectures with a motivation to improve learning the energy function in an unsupervised setting. The thesis is organized as follows. First, we list out a number of common strategies to shape a good energy function by learning. Among these, we mainly target at two strategies and extend the frontier of them. The resulted models from this thesis demonstrate several applications made possible by using no or few labeled data. It covers a wide spectrum of computer vision and language tasks, such as generation, text summarization, text style-transfer and transfer/semi-supervised learning. |
| 일반주제명 | Artificial intelligence. |
| 언어 | 영어 |
| 바로가기 | 
: 이 자료의 원문은 한국교육학술정보원에서 제공합니다. |
