Complex number deep neural networks applied to natural language processing

October 4, 2018
Versione stampabile

Date and Time: October 4, 2018 - 11.30 am
Venue: Ferrari 1 Building, via Sommarive 5 - Povo - Room A212

Speaker

  • Prof. Renato De Mori - emeritus professor at Mc Gill University (Canada) and at the University of Avignon (France)

Abstract

Complex number multilayer perceptron models have been proposed in the past mostly for applications in image processing and robot arm control. Recently, new quaternion neural networks have been developed for encoding multiple views of the same feature such as an image pixel color or the speech energy evolution in a given frequency band. In this way, a single quaternion of weights is multiplied by a single quaternion of inputs with a substantial reduction in the number of parameters to estimate, an advantage for applications with limited resources. 

Quaternion numbers and their algebra will be briefly introduced. New denoising autoencoders and quaternion convolutional neural networks (QCNN) will be described and results on conversation analysis will be presented. Deep quaternion valued recurrent neural networks (QRNN) will also be introduced.  Applications of QCNNs and QRNNs to end-to-end automatic speech recognition will be discussed showing significant improvements and architecture complexity reduction with respect to corresponding solutions with real valued model parameters. Future work on the introduction of mechanisms of attention and various types of memories will be briefly outlined.

About the Speaker

Renato De Mori is emeritus professor at Mc Gill University (Canada) and at the University of Avignon (France). He has been full professor in Italy (University of Turin), Canada and France. He is a Life Fellow of the Computer Society and has been distinguished lecturer of the Signal Processing Society of the Institute of Electrical and Electronic Engineers (IEEE). He has been scientific leader of projects in Canada (The Institute for Robotics and Intelligent Systems, a Canadian Network of Centers of Excellence) and in Europe (Language Understanding for telephone Applications ( LUNA) and is actually working on the m2cr CHIST ERA European  project in cooperation with the MILA laboratory in Montreal.

He is actually member of the Computational Linguistic group at Mc Gill University and the speech communication group of the Avignon Computer Science Laboratory.

He has been Chief Editor of Speech Communication (2003-2005) (member of the editorial board since 1982), associate Editor of IEEE Transactions on Pattern Analysis and Machine Intelligence (1998-1992) and of the IEEE Transactions on Audio, Speech and Language Processing (2009-2014). He is in the Editorial board of Computer Speech and Language (since 1988) He is author/editor of four books and more than 100 publications in international journals and conferences. 

 He has been Vice-President, Research Centre de recherche informatique de Montréal, Member then Chairman Natural Science and Engineering Research Council of Canada (NSERC) Computing and Information Sciences Grant Section Committee,.He has been member of Executive Advisory Board IBM Canada Toronto Research Lab, IEEE Speech technical committee, Scientific  Council IM2 – Swiss Network of excellence, Canadian Chairs Review panel, Interdisciplinary Adjudication Committee (NSERC), Interdisciplinary Committee Canadian Foundation for Innovation, NSERC Strategic Grant Committee Communication and Computers Panel, Advisory Board Information Systems National Research Council of Canada, Scientific Council Centre National  d'Etudes des Télécommunications, Lannion, France, Scientific Council CSELT Telecom Italy. 

His research interest has been in spoken language processing, computer architecture, and software engineering. He is actually working on complex number deep neural networks with application to automatic speech recognition and the extraction of semantic information from spoken documents.

Contact: giuseppe.riccardi [at] unitn.it (Giuseppe Riccardi)