Find results with:

Add another field

Search by date

to

Searching collections:

Dartmouth Dissertations

Add or remove collections

Home Dartmouth Dissertations fang2016 1

Reference URL

To link to the entire object, paste this link in email, IM or document

To embed the entire object, paste this HTML in website

To link to this page, paste this link in email, IM or document

To embed this page, paste this HTML in website

Representation learning beyond labels / by Chen Fang

you wish to report:

Your comment:

Your Name:

Submit

Cancel

...

Back to top

Powered by CONTENTdm®

Author	Fang, Chen
Title	Representation learning beyond labels / by Chen Fang
Date	2016
Abstract	The performance of machine learning algorithms largely depends on data representation. As a critical step in machine learning, representation learning (feature learning) learns a transformation of training data to give a new representation that can be utilized for machine learning tasks. Supervised representation learning methods typically rely on labeled datasets for supervision knowledge. However, in many settings labeled data has limitations, such as sample size, bias and weak labels. This thesis builds on the observation that supervision knowledge can be found in various forms beyond labels. Therefore, we consider the problem of designing representation learning methods that do not rely exclusively on explicit labels, for important prediction tasks. First, in order to overcome the limitation of dataset size, we develop a multi-task representation learning algorithm for link prediction on network data that leverages not only link labels but implicit correlation between tasks. The algorithm is based on metric learning, and capable of joint learning across multiple different but related networks for improved predictive performance. Second, to reduce the negative effect of data bias on image recognition, we learn domain-invariant image representations via metric learning. By exploiting for supervision the explicit class labels as well as the domain information associated with training data, the learned image representations can generalize better to new domains under the condition of limited and biased training data. Finally, we present a flexible framework that can learn visual features from online visual data and user behavior data (e.g., social media websites and their user view history, image search engine and user clicks). The proposed framework is applied to two large-scale visual datasets collected from real-world Internet scenarios. In both, competitive and generalizable image features are learned.
Description	1 online resource (xii, 134 leaves) : illustrations (mostly color)
Format	application/pdf
College	Dartmouth College -- Department of Computer Science.
Issuing Body	Dartmouth Dissertations are made available through the Dartmouth Digital Library.
Degree	Thesis (Ph.D)--Dartmouth College.
Copyright	Made available by the author under a Creative Commons Attribution-NonCommercial License.; Made available by the Trustees of Dartmouth College under a Creative Commons Attribution-NonCommercial License. https://creativecommons.org/licenses/by-nc/4.0/
DOI	https://doi.org/10.1349/ddlp.1959

Title	fang2016 1
Full Text	REPRESENTATION LEARNING BEYOND LABELS A Thesis Submitted to the Faculty in partial ful llment of the requirements for the degree of Doctor of Philosophy in Computer Science by Chen Fang DARTMOUTH COLLEGE Hanover, New Hampshire March 1, 2016 Examining Committee: (chair)Daniel Rockmore, Ph.D. Hany Farid, Ph.D. Devin Balkcom, Ph.D. Hailin Jin, Ph.D. F. Jon Kull, Ph.D. Dean of Graduate Studies

	All
	Pages Enter page numbers or page ranges separated by commas. For example: 1,3,5-12

Find results with:

Searching collections:

Home Dartmouth Dissertations fang2016 1

Representation learning beyond labels / by Chen Fang

fang2016 1

Object Description

Description

Representation learning beyond labels / by Chen Fang

Representation learning beyond labels / by Chen Fang