Events Calendar

Cognitive Motivations for Non-negative Matrix Factorizations

Speaker: Professor Hugo Van hamme, Dept. of Electrical Engineering (ESAT), Katholieke Universiteit Leuven, Belgium

Talk Title: Cognitive Motivations for Non-negative Matrix Factorizations

Abstract: Non-negative Matrix Factorization (NMF) and related latent variable methods such as Latent Dirichlet Allocation have been applied in many fields of engineering such as speech, text, and image processing to discover relations with great success. Its core capability is to decompose wholes (scenes) into parts represented in the matrix factors. In their 1999 Nature paper, Lee and Seung point out some resemblances between non-negative matrix factorization and the brain. For one, like neural firing rates, NMF assumes non-negative quantities, which lead to sparse representations. In this talk, additional similarities will be discussed:
- NMF can be viewed as a neural network with an intrinsic lateral inhibition mechanism,
- the matrix factors can be obtained using operations that can be implemented in neurons,
- NMF can learn with, without, or with weak cross-modal supervision,
- learning can be made incremental,
- NMF can explain time perception with integrate-and-fire neurons.
Latent variable methods should hence not be seen purely as statistical inference problems, but can be motivated from a cognitive perspective.

Biography: Prof. Hugo Van hamme received the masters degree in electomechanical engineering from Vrije Universiteit Brussel, Belgium in 1987, the masters degree in controls systems from Imperial College, U.K. in 1988 and the Ph.D. in electrical engineering from Vrije Universiteit Brussel in 1992. In 1993, he joined Lernout & Hauspie n.v. and held positions of senior researcher, team leader, director, and senior director of research. In 2001, he joined ScanSoft as senior director of research and engineering for automotive and embedded products. In 2002, he was appointed professor at the Department of Electrical Engineering of KU Leuven where he teaches courses in speech processing and algebra. His current research interests are robust automatic speech recognition, vocabulary learning, technology for speech therapy, and audio analysis. He is the author of over 150 publications.

Host: Dr. Maarten Van Segbroeck and Professor Shrikanth Narayanan

Location: Ronald Tutor Hall of Engineering (RTH) - 320

Audiences: Everyone Is Invited

Contact: Mary Francis