Graph learning: perspectives in statistics and signal processing
Event details
| Date | 05.08.2016 |
| Hour | 10:30 › 12:30 |
| Speaker | Rodrigo Cerqueira Gonzalez Pena |
| Location | |
| Category | Conferences - Seminars |
EDIC Candidacy Exam
Exam President: Prof. Jean-Philippe Thiran
Thesis Director: Prof. Pierre Vandergheynst
Co-examiner: Prof. Dimitri Van De Ville
Background papers:
Model selection through sparse maximum likelihood estimation for multivariate Gaussian or binary data, by O. Banerjee, L. El Ghaoui, and A. d’Aspremont.
How to learn a graph from smooth signals, by V. Kalofolias.
Compressive spectral clustering, by N. Tremblay, G. Puy, R. Gribonval, P. Vandergheynst
Abstract:
It is increasingly common for machine learning algorithms to take advantage of the relationship between data points to improve on their classification or regression performance. This data relationship can be efficiently modelled as a graph, and it is known that the performance of algorithms such as spectral clustering, or label propagation depend solely on the quality of this graph. Nonetheless, if this graphical structure is not know in advance, then it should be learned from the data observations. We present two approaches to solving this problem: the first learns a graph in which the signals given as input are smooth according to some objective measure; the second learns a graphical representation in the context of a Gauss-Markov Random Field (GMRF) model, by maximising the likelihood of the model conditional on the observed data. We finally present a breadth paper, highlighting the importance of the learned graph for the problem of compressive spectral clustering.
Exam President: Prof. Jean-Philippe Thiran
Thesis Director: Prof. Pierre Vandergheynst
Co-examiner: Prof. Dimitri Van De Ville
Background papers:
Model selection through sparse maximum likelihood estimation for multivariate Gaussian or binary data, by O. Banerjee, L. El Ghaoui, and A. d’Aspremont.
How to learn a graph from smooth signals, by V. Kalofolias.
Compressive spectral clustering, by N. Tremblay, G. Puy, R. Gribonval, P. Vandergheynst
Abstract:
It is increasingly common for machine learning algorithms to take advantage of the relationship between data points to improve on their classification or regression performance. This data relationship can be efficiently modelled as a graph, and it is known that the performance of algorithms such as spectral clustering, or label propagation depend solely on the quality of this graph. Nonetheless, if this graphical structure is not know in advance, then it should be learned from the data observations. We present two approaches to solving this problem: the first learns a graph in which the signals given as input are smooth according to some objective measure; the second learns a graphical representation in the context of a Gauss-Markov Random Field (GMRF) model, by maximising the likelihood of the model conditional on the observed data. We finally present a breadth paper, highlighting the importance of the learned graph for the problem of compressive spectral clustering.
Practical information
- General public
- Free
Contact
- Cecilia Chapuis EDIC