Vai al contenuto principale
Oggetto:

Exact-Approximate Bayesian Inference for Gaussian Processes

Oggetto:

Exact-Approximate Bayesian Inference for Gaussian Processes

Oggetto:

Academic year 2014/2015

Course ID
SEM-ABIGP
Year
1° anno 2° anno 3° anno
Teaching period
Seminario
Type
Seminario
Course disciplinary sector (SSD)
INF/01 - informatica
Delivery
Tradizionale
Language
Inglese
Attendance
Facoltativa
Type of examination
Non prevista
Oggetto:

Sommario del corso

Oggetto:

Program

Hierarchical models are commonly employed to model complex phenomena. The importance of achieving a sound quantification of uncertainty in predictions by characterizing the posterior distribution over all parameters exactly has been demonstrated in several applications. In this talk, I will focus on the Bayesian treatment of a class of hierarchical models involving Gaussian Processes (GP) using Markov chain Monte Carlo (MCMC) methods. After discussing why MCMC is the only way to infer all parameters exactly in GP models and pointing out the challenges in doing so, I will present the so called Exact-Approximate MCMC approach that offers a practical solution to this problem by overcoming the huge challenges associated with the exact computation of the intractable marginal likelihood of such models. In particular, the Exact-Approximate MCMC approach yields samples from the correct posterior distribution over model parameters, but only requires an unbiased estimate of the intractable marginal likelihood. I will finally present ways to construct unbiased estimates of the marginal likelihood in GP models, and conclude the talk by presenting results on several benchmark data sets and on a multi-class multiple-kernel classification problem on neuroimaging data.

Suggested readings and bibliography



Oggetto:

Class schedule

DaysTimeClassroom
Venerdì15:15 - 17:15

Lessons: dal 31/01/2014 to 31/01/2014

Notes: Sala seminari, Dipartimento di "Economia e Statistica", Campus Luigi Einaudi, Lungo Dora Siena 100A, 3 piano.

Oggetto:

Note

The seminar will be held by

Maurizio Filippone
School of Computing Science - University of Glasgow

Oggetto:
Last update: 27/01/2014 11:20
Location: https://dott-informatica.campusnet.unito.it/robots.html
Non cliccare qui!