Grant: Discovery Grants Program - Individual
The University of British Columbia
$37,000.00
Mathematical problems in analysis, processing, and use of deep neural networks in signal processing
EF 2024-25
Résumé
Obtaining effective signal models has been a central problem in applied mathematics, mathematical signal processing, and information theory for decades. Accurate signal models allow the design of mathematically sound and practicable methods to acquire information about signals (as in Nyquist-Shannon sampling theory or compressed sensing), process signals (e.g., denoising), compress signals (as in the case JPEG), and transmit them effectively over noisy channels. As such, results improving the signal models and enabling faster and more effective methods are crucial to address many technological challenges we face like efficient analog-to-digital conversion, compression, as well as inverse problems like medical and seismic imaging, inpainting, and phase retrieval (e.g., in x-ray crystallography).
Classical approaches often model signal classes as specific function spaces (e.g., Sobolev, bandlimited functions, functions with bounded variation). This functional analytic paradigm, together with techniques using mathematical tools from harmonic analysis, probability theory, discrete math, number theory, and more, has been instrumental in developing successful methods for sampling, quantization, compression and inverse problems. Yet, it is also well-known that the function space models are ad hoc and overly expressive, and thus they limit the success of resulting applications. Recently, a new alternative paradigm to modeling signals has been proposed in the literature. This paradigm relies on generative models that are learned from data by training a deep neural network, referred as "Deep Generative Models" (DGMs). DGMs enable sampling the prior signal distribution by sampling a lower dimensional latent space. For example, one can train a DGM that produces images of faces (not belonging to any real person) when fed low-dimensional random input.
My research program seeks to develop theory and algorithms that exploit the interplay between DGMs and the rich suite of techniques relying on classical methods. The proposed research will build on my established track record on the theory of (quantized) compressed sensing and related techniques based on work during the period of my current and past Discovery Grants, together with my recent work on using DGMs in solving certain inverse problems. Specific directions of focus include devising theory and algorithms for various classes of inverse problems using generative priors, developing signal sampling and quantization theory for DGMs (obtaining digital representations for signals), and developing a quantization theory for deep neural networks (digitizing network parameters). These will provide HQP at all levels with the opportunity to gain expertise in hot research areas where there are many open problems but also where their skills will be aligned with the needs of the employment market, be it academic or non-academic (e.g., industry, public sector, and NGOs).
Programme
Discovery Grants Program - Individual
Bénéficiaire
Emplacement
British Columbia, Canada
Type de bénéficiaire
Academia
Type
Grant
Source
NSERC's Awards Data
Tel que publié
Application id
RGPIN-2023-04532
Name nom
Yilmaz, Ozgur
Department département
Mathematics
Organization id
2
Institution établissement
The University of British Columbia
Province en
British Columbia
Province fr
Colombie-Britannique
Country en
CANADA
Country fr
CANADA
Fiscal year exercice financier
2024
Competition year année de concours
2023
Award amount
37000
Program id
RGPIN
Program name en
Discovery Grants Program - Individual
Program name fr
Programme de subventions à la découverte - individuelles
Group en
DISCOVERY RESEARCH
Group fr
RECHERCHE AXÉE SUR LA DÉCOUVERTE
Committee code
1508
Committee name en
Mathematics and Statistics
Committee name fr
Mathématiques et statistique
Area of application code
1205
Area of application group en
Advancement of knowledge
Area of application group fr
Avancement des connaissances
Area of application en
Mathematical sciences
Area of application fr
Sciences mathématiques
Research subject code
2959
Research subject group en
Applied mathematics
Research subject group fr
Mathématiques appliquées
Research subject en
Mathematics of communications
Research subject fr
Mathématiques des télécommunications
Field of research list names en
Not available
Field of research list names fr
Non disponible
Socioeconomic objective list names en
Not available
Socioeconomic objective list names fr
Non disponible
Application title
Mathematical problems in analysis, processing, and use of deep neural networks in signal processing
Keywords
compressed sensing; quantization; inverse problems; compression; denoising; signal modeling; deep generative models
Application summary
Obtaining effective signal models has been a central problem in applied mathematics, mathematical signal processing, and information theory for decades. Accurate signal models allow the design of mathematically sound and practicable methods to acquire information about signals (as in Nyquist-Shannon sampling theory or compressed sensing), process signals (e.g., denoising), compress signals (as in the case JPEG), and transmit them effectively over noisy channels. As such, results improving the signal models and enabling faster and more effective methods are crucial to address many technological challenges we face like efficient analog-to-digital conversion, compression, as well as inverse problems like medical and seismic imaging, inpainting, and phase retrieval (e.g., in x-ray crystallography).
Classical approaches often model signal classes as specific function spaces (e.g., Sobolev, bandlimited functions, functions with bounded variation). This functional analytic paradigm, together with techniques using mathematical tools from harmonic analysis, probability theory, discrete math, number theory, and more, has been instrumental in developing successful methods for sampling, quantization, compression and inverse problems. Yet, it is also well-known that the function space models are ad hoc and overly expressive, and thus they limit the success of resulting applications. Recently, a new alternative paradigm to modeling signals has been proposed in the literature. This paradigm relies on generative models that are learned from data by training a deep neural network, referred as "Deep Generative Models" (DGMs). DGMs enable sampling the prior signal distribution by sampling a lower dimensional latent space. For example, one can train a DGM that produces images of faces (not belonging to any real person) when fed low-dimensional random input.
My research program seeks to develop theory and algorithms that exploit the interplay between DGMs and the rich suite of techniques relying on classical methods. The proposed research will build on my established track record on the theory of (quantized) compressed sensing and related techniques based on work during the period of my current and past Discovery Grants, together with my recent work on using DGMs in solving certain inverse problems. Specific directions of focus include devising theory and algorithms for various classes of inverse problems using generative priors, developing signal sampling and quantization theory for DGMs (obtaining digital representations for signals), and developing a quantization theory for deep neural networks (digitizing network parameters). These will provide HQP at all levels with the opportunity to gain expertise in hot research areas where there are many open problems but also where their skills will be aligned with the needs of the employment market, be it academic or non-academic (e.g., industry, public sector, and NGOs).