TOPIC Difficulties & Solutions of Bayesian Inferences in Solving Inverse Problems of Electromagnetic Scattering 
AREA Microwave & RF 
SPEAKER Dr Caifang Cai, Laboratory of Signals & Systems (L2S),CNRS-SUPELEC-PARIS SUD  
DATE 5 January 2015, Monday 
TIME 3:00 pm to 4:00 pm 
VENUE E3-06-04, Engineering Blk 3, Faculty of Engineering, NUS 
FEES No Charge 

SYNOPSIS
In solving Electromagnetic (EM) scattering problems, Bayesian inferences have great potentials in noise robustness, parameter inversion accuracy and the ability of managing variant prior information. However, they often suffer from extremely high computational cost. This is mainly due to the fact that the forward models that we are dealing with are often non-linear and computationally expensive.
To overcome such a difficulty in computational cost, in this talk, we are first going to present a metamodeling method which is a surrogate forward modeling method based on pre-training of databases by using simulation softwares. This method transfers the high computational cost from the inverse problem to the pre-training of metamodel databases. By referring to this, a forward projection which previously costs minutes on a standard computer can be done within less a second. This considerably reduces the computational cost of the Bayesian inferences and makes them usable in practical applications. As for the Bayesian inferences, in this talk, we are going to discuss two of them specifically. The first is a Markov Chain Monte-Carlo (MCMC) sampling method proposed for parameter inversion; the second is a Nested Sampling (NS) proposed for automatic model selection. The MCMC method allows us to estimate the unknown parameters while the NS method makes it possible for us the tell the correct object model.
Against simulations and laboratory controlled experiments, we validate these two methods. The results confirm their high efficiency in model selection and parameter inversion. Yet, the results also show that the computational cost increases exponentially as a function of the number of unknown parameters. This indicates that attention must be paid to the parametrization of problems, a small number of unknown parameters having to be preferred. It also illustrates the strong need of improvements in the metamodeling method in the future in order to cope with problems with large numbers of parameters.

 

ABOUT THE SPEAKER
In solving Electromagnetic (EM) scattering problems, Bayesian inferences have great potentials in noise robustness, parameter inversion accuracy and the ability of managing variant prior information. However, they often suffer from extremely high computational cost. This is mainly due to the fact that the forward models that we are dealing with are often non-linear and computationally expensive.
To overcome such a difficulty in computational cost, in this talk, we are first going to present a metamodeling method which is a surrogate forward modeling method based on pre-training of databases by using simulation softwares. This method transfers the high computational cost from the inverse problem to the pre-training of metamodel databases. By referring to this, a forward projection which previously costs minutes on a standard computer can be done within less a second. This considerably reduces the computational cost of the Bayesian inferences and makes them usable in practical applications. As for the Bayesian inferences, in this talk, we are going to discuss two of them specifically. The first is a Markov Chain Monte-Carlo (MCMC) sampling method proposed for parameter inversion; the second is a Nested Sampling (NS) proposed for automatic model selection. The MCMC method allows us to estimate the unknown parameters while the NS method makes it possible for us the tell the correct object model.
Against simulations and laboratory controlled experiments, we validate these two methods. The results confirm their high efficiency in model selection and parameter inversion. Yet, the results also show that the computational cost increases exponentially as a function of the number of unknown parameters. This indicates that attention must be paid to the parametrization of problems, a small number of unknown parameters having to be preferred. It also illustrates the strong need of improvements in the metamodeling method in the future in order to cope with problems with large numbers of parameters.

 

REMARKS, IF ANY
Jointly organized by IEEE Singapore MTT/AP Chapter and Dept of Electrical & Computer Engineering, NUS 

 | Back to seminar list |

All are welcome to attend these Seminars. Should you have any enquiries, please contact Ms. Lily Png at 6516 6509 or you can e-mail to Ms Lily Png ().