Grant: Strategic Projects - Group
Université de Montréal
$127,304.00
Guided Learning of Representations for Web Data
EF 2010-11
Résumé
Web services such as search engines rely on the interpretation of the content of Web pages: text, images, video. Those services could be drastically improved if we could improve the ability of computers to disambiguate words in their context, to interpret images in Web pages, or to correctly associate a document and an image. To analyze and compare words, sentences, documents, images and video, while better taking into account their meaning, their raw digital forms can be transformed into new representations. Many machine learning algorithms have been developed around the world to learn such representations. Among these algorithms, much attention has recently been devoted to so-called deep architectures, which decompose the problem into a series of transformations, from low-level features to higher-level features. The proposed research program focusses on overcoming current limitations of these algorithms and on taking advantage of huge datasets. Such datasets may be those currently available from the Web, or be automatically synthetized so as to illustrate different levels of abstraction about text and images not available explicitly from Web data. It is proposed to guide the training of machines to capture these abstractions. This will be achieved by taking advantage of the fact that higher-level features tend to remain for many consecutive sentences or images, and also taking inspiration from human and animal learning to order examples in a structured curriculum.
Programme
Strategic Projects - Group
Bénéficiaire
Emplacement
Quebec, Canada
Type de bénéficiaire
Academia
Type
Grant
Source
NSERC's Awards Data
Tel que publié
Application id
381312-2009
Name nom
Bengio, Yoshua
Department département
Informatique et recherche opérationnelle
Organization id
63
Institution établissement
Université de Montréal
Province en
Québec
Province fr
Québec
Country en
CANADA
Country fr
CANADA
Fiscal year exercice financier
2010
Competition year année de concours
2009
Award amount
127304
Program id
STPGP
Program name en
Strategic Projects - Group
Program name fr
Projets stratégiques - De groupe
Group en
Research partnerships programs
Group fr
Programmes de partenariat
Committee code
1041
Committee name en
Strategic Projects Selection Panel - Communications and Information
Committee name fr
Comité de sélection des projets stratégiques - Communication et information
Area of application code
1207
Area of application group en
Advancement of knowledge
Area of application group fr
Avancement des connaissances
Area of application en
Information, computer and communication technologies
Area of application fr
Technologie de l'information, des ordinateurs et des réseaux de communications
Research subject code
2805
Research subject group en
Artificial intelligence
Research subject group fr
Intelligence artificielle
Research subject en
Learning and inference theories
Research subject fr
Théories de l'apprentissage et l'inférence
Partie
2
Nb partie
3
Application title
Guided Learning of Representations for Web Data
Keyword
NSERC
Application summary
Web services such as search engines rely on the interpretation of the content of Web pages: text, images, video. Those services could be drastically improved if we could improve the ability of computers to disambiguate words in their context, to interpret images in Web pages, or to correctly associate a document and an image. To analyze and compare words, sentences, documents, images and video, while better taking into account their meaning, their raw digital forms can be transformed into new representations. Many machine learning algorithms have been developed around the world to learn such representations. Among these algorithms, much attention has recently been devoted to so-called deep architectures, which decompose the problem into a series of transformations, from low-level features to higher-level features. The proposed research program focusses on overcoming current limitations of these algorithms and on taking advantage of huge datasets. Such datasets may be those currently available from the Web, or be automatically synthetized so as to illustrate different levels of abstraction about text and images not available explicitly from Web data. It is proposed to guide the training of machines to capture these abstractions. This will be achieved by taking advantage of the fact that higher-level features tend to remain for many consecutive sentences or images, and also taking inspiration from human and animal learning to order examples in a structured curriculum.