November 08, 2018
14:00 - 14:45
VAES ROOM, LLN (blind side)
Research Seminar by Masashi SHIMBO, NARA Institute of Technology, Japan
Reverse linear regression for cross-domain retrieval
We discuss the use of linear regression in cross-domain retrieval and zero-shot learning. Traditionally for these tasks, a mapping from the space of query objects to that of target objects was sought, in the simplest case, with linear regression techniques such as OLS or ridge regression. Then, using the learned mapping, query objects are mapped to the target space wherein nearest target objects are retrieved.
Here we propose the opposite, that is, we learn to map target objects in the query space, and the nearest neighbor retrieval is carried in the query space.
This "reverse" mapping effectively suppresses the emergence of "hubs"
in the subsequent nearest neighbor retrieval step, which contributes the improved accuracy. Assuming a simple data model, we show that the proposed approach reduces hubness whereas the traditional "forward mapping" approach actually promotes hubness. Empirical evaluation also shows that our method performs better empirically in the tasks of bilingual lexicon extraction and image labeling.
Masashi Shimbo, Nara Institute of Technology, Japan (This is a joint work with Yutaro Shigeto, Ikumi Suzuki, and Kazuo Hara)