CeSIP Seminar Series

The CeSIP Seminar Series is held on a fortnightly basis on Thursdays from 12-1 pm in R3.19 of McCance Building. Talks are presented by CeSIP academic staff, PhD students and special visitors. All are welcome to attend.

If you would like to join the seminar's mailing list please send an e-mail to Carmine Clemente at carmine.clemente@strath.ac.uk with 'CeSIP Seminar' in the subject line.

Upcoming Seminars
Stephan Weiss 16 May 2013 Bayesian Belief Propagation Bayesian belief networks are based on directed acyclic graphs which can model the internal dependencies between variables found in many estimation problems. Two particular uses appear to be in scenarios where variables are random and characterised by their distributions, and cases where the independence of some branchesof the network, as established by the network model, enables distributed processing. The aim is to provide some insight into the background of Bayes, show how Bayesian belief networks operate, how they can be implemented, and some potential application.
Jaime Zabalza 23 May 2013 Combination of Support Vector Machine and Principal Component Analysis for classification tasks Combination of Principal Component Analysis (PCA) and Support Vector Machine (SVM) is a powerful tandem for classification tasks in many applications, from hyperspectral imaging to radar signal classification. In this seminar, some useful experiences and tips related to the implementation of these two techniques are provided. PCA is a widely used technique to convert high dimensional data to linearly uncorrelated components. However, PCA usually suffers from extremely large dimension in calculating the covariance matrix, therefore, different implementations from the conventional one can be applied in order to improve computational aspects. A proposed approach here lays in obtaining partial covariance matrices separately and constructing the overall covariance matrix by the addition of these partial elements, which can significantly reduce the scale of required memory, avoiding computational difficulties. On the other hand, SVM is nowadays one of the best classifiers for data classification and prediction. A brief review is summarised to cover its implementation and practical development, highlighting the importance of an appropriate configuration and parameters selection in practical applications.

Upcoming Seminars

Stephan Weiss

16 May 2013

Bayesian Belief Propagation

Bayesian belief networks are based on directed acyclic graphs which can model the internal dependencies between variables found in many estimation problems. Two particular uses appear to be in scenarios where variables are random and characterised by their distributions, and cases where the independence of some branchesof the network, as established by the network model, enables distributed processing.

The aim is to provide some insight into the background of Bayes, show how Bayesian belief networks operate, how they can be implemented, and some potential application.

Jaime Zabalza

23 May 2013

Combination of Support Vector Machine and Principal Component Analysis for classification tasks

Combination of Principal Component Analysis (PCA) and Support Vector Machine (SVM) is a powerful tandem for classification tasks in many applications, from hyperspectral imaging to radar signal classification. In this seminar, some useful experiences and tips related to the implementation of these two techniques are provided.

PCA is a widely used technique to convert high dimensional data to linearly uncorrelated components. However, PCA usually suffers from extremely large dimension in calculating the covariance matrix, therefore, different implementations from the conventional one can be applied in order to improve computational aspects. A proposed approach here lays in obtaining partial covariance matrices separately and constructing the overall covariance matrix by the addition of these partial elements, which can significantly reduce the scale of required memory, avoiding computational difficulties.

On the other hand, SVM is nowadays one of the best classifiers for data classification and prediction. A brief review is summarised to cover its implementation and practical development, highlighting the importance of an appropriate configuration and parameters selection in practical applications.