[cs-talks] Kun He, PhD Thesis Proposal, 10/17 @11am, MCS 148
Harrington, Jacob Walter
jwharrin at bu.edu
Fri Oct 13 16:37:58 EDT 2017
Learning Deep Embeddings by Learning to Rank
Kun He, PhD Student, Image and Video Computing, Boston University Department of Computer Science
Tuesday, October 17th from 11am – 12:30pm in 111 Cummington Mall, Room 148
We study the problem of learning low-dimensional vector embeddings of high-dimensional data. This is an important problem in many application areas such as computer vision, as embedding techniques can both perform dimensionality reduction and capture semantic similarities. In this thesis, we propose to learn vector embeddings using deep neural networks, and we propose learning to rank formulations that optimize the embeddings for nearest neighbor retrieval. For learning binary embeddings, also known as supervised hashing, we develop a gradient-based optimization framework, and propose two solutions based on optimizing novel listwise ranking losses. In our feasibility study, these solutions establish the state-of-the-art in both 1) supervised hashing for image retrieval, and 2) learning feature descriptors for local image patches. Remaining work include extending our techniques to address the problem of learning real-valued vector embeddings, as well as applying them in the task of cross-modal retrieval between language and video.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the cs-talks