On Deep Multi-View Representation Learning

Weiran Wang, Raman Arora, Karen Livescu, Jeff Bilmes
Proceedings of the 32nd International Conference on Machine Learning, PMLR 37:1083-1092, 2015.

Abstract

We consider learning representations (features) in the setting in which we have access to multiple unlabeled views of the data for representation learning while only one view is available at test time. Previous work on this problem has proposed several techniques based on deep neural networks, typically involving either autoencoder-like networks with a reconstruction objective or paired feedforward networks with a correlation-based objective. We analyze several techniques based on prior work, as well as new variants, and compare them experimentally on visual, speech, and language domains. To our knowledge this is the first head-to-head comparison of a variety of such techniques on multiple tasks. We find an advantage for correlation-based representation learning, while the best results on most tasks are obtained with our new variant, deep canonically correlated autoencoders (DCCAE).

Cite this Paper

BibTeX
@InProceedings{pmlr-v37-wangb15, title = {On Deep Multi-View Representation Learning}, author = {Wang, Weiran and Arora, Raman and Livescu, Karen and Bilmes, Jeff}, booktitle = {Proceedings of the 32nd International Conference on Machine Learning}, pages = {1083--1092}, year = {2015}, editor = {Bach, Francis and Blei, David}, volume = {37}, series = {Proceedings of Machine Learning Research}, address = {Lille, France}, month = {07--09 Jul}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v37/wangb15.pdf}, url = {https://proceedings.mlr.press/v37/wangb15.html}, abstract = {We consider learning representations (features) in the setting in which we have access to multiple unlabeled views of the data for representation learning while only one view is available at test time. Previous work on this problem has proposed several techniques based on deep neural networks, typically involving either autoencoder-like networks with a reconstruction objective or paired feedforward networks with a correlation-based objective. We analyze several techniques based on prior work, as well as new variants, and compare them experimentally on visual, speech, and language domains. To our knowledge this is the first head-to-head comparison of a variety of such techniques on multiple tasks. We find an advantage for correlation-based representation learning, while the best results on most tasks are obtained with our new variant, deep canonically correlated autoencoders (DCCAE).} }
Endnote
%0 Conference Paper %T On Deep Multi-View Representation Learning %A Weiran Wang %A Raman Arora %A Karen Livescu %A Jeff Bilmes %B Proceedings of the 32nd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2015 %E Francis Bach %E David Blei %F pmlr-v37-wangb15 %I PMLR %P 1083--1092 %U https://proceedings.mlr.press/v37/wangb15.html %V 37 %X We consider learning representations (features) in the setting in which we have access to multiple unlabeled views of the data for representation learning while only one view is available at test time. Previous work on this problem has proposed several techniques based on deep neural networks, typically involving either autoencoder-like networks with a reconstruction objective or paired feedforward networks with a correlation-based objective. We analyze several techniques based on prior work, as well as new variants, and compare them experimentally on visual, speech, and language domains. To our knowledge this is the first head-to-head comparison of a variety of such techniques on multiple tasks. We find an advantage for correlation-based representation learning, while the best results on most tasks are obtained with our new variant, deep canonically correlated autoencoders (DCCAE).
RIS
TY - CPAPER TI - On Deep Multi-View Representation Learning AU - Weiran Wang AU - Raman Arora AU - Karen Livescu AU - Jeff Bilmes BT - Proceedings of the 32nd International Conference on Machine Learning DA - 2015/06/01 ED - Francis Bach ED - David Blei ID - pmlr-v37-wangb15 PB - PMLR DP - Proceedings of Machine Learning Research VL - 37 SP - 1083 EP - 1092 L1 - http://proceedings.mlr.press/v37/wangb15.pdf UR - https://proceedings.mlr.press/v37/wangb15.html AB - We consider learning representations (features) in the setting in which we have access to multiple unlabeled views of the data for representation learning while only one view is available at test time. Previous work on this problem has proposed several techniques based on deep neural networks, typically involving either autoencoder-like networks with a reconstruction objective or paired feedforward networks with a correlation-based objective. We analyze several techniques based on prior work, as well as new variants, and compare them experimentally on visual, speech, and language domains. To our knowledge this is the first head-to-head comparison of a variety of such techniques on multiple tasks. We find an advantage for correlation-based representation learning, while the best results on most tasks are obtained with our new variant, deep canonically correlated autoencoders (DCCAE). ER -
APA
Wang, W., Arora, R., Livescu, K. & Bilmes, J.. (2015). On Deep Multi-View Representation Learning. Proceedings of the 32nd International Conference on Machine Learning, in Proceedings of Machine Learning Research 37:1083-1092 Available from https://proceedings.mlr.press/v37/wangb15.html.

Related Material