• Versión en Español  
• General information
• Bio and CV
• Research
  • Research lines
  • Publications
  • Software
• Courses
  • Open Courses (OCW)
  • UC3M Courses
• Contact
  • Send e-mail to Marcelino Lázaro

Research

 

Research lines

In the general framework of digital signal processing and statistical detection theory, the current research lines are:

• Classification in singular problems

Statistical machine learning methods are used to solve imbalanced classification problem as well as example dependent cost problems.
A classification problem is said to be imbalanced when the number of available patterns for the possible classes is different. In this kind of problems the probability of error tends to be higher for the minoritary classes, because they can be shadowed by the higher number of examples of the majoritary classes. For this reason, conventional classification methods designed for balanced classification problems provide in general poor results in these problems.

An example dependent cost classification problem is a problem were the cost of a misclassification depends on each pattern. Again, conventional methods for classification usually provide poor results for these problems.

There exist several alternatives to deal with the imbalanceness or with the example dependency. The alternative that has been considered consists in designing an appropriate cost function. In particular, the Bayesian statistical decision theory provides the mechanism to consider the probabilities of the different classes as well as the cost associated to every possible decision, thus providing a suitable framework to face this problem.

• Ordinal classification

In this kind of classification problems the possible classes are sorted in an specific arrangement, which means that the cost of a misclassification depends on the distance between the right and wrong classes in the specified arrangement.

As in the previous research line, this problem has been faced by defining appropriate cost functions in the framework of the Bayesian statistical detection theory.

• Sensor networks

Development of detection and localization rules, and performance analysis, in distributed sensor networks. Specially interesting is the case of networks where the available lectures of the sensing nodes are not conditionally independent, a realistic situation in many applications that makes more complicated the design and analysis of the networks.

Some previous research lines:

• Receivers for communications

Design of receivers for digital communication systems, in particular channel equalizers.

The channel equalization problem has been faced using the Infomation Theory framework, by using quantitative information measures or several divergence measures in the design of cost functions that are appropriate for blind channel equalization.

• Simultaneous recostruction of a function and its derivatives
  

In some applications where a given function has to be reconstructed from samples it is of special importance not only to have an accurate reconstruction of the function itself but also of its derivatives. An example is the nonlinear device modeling if the effect of the intermodulation distortion has to be considered.

Several tools have been proposed for the simultaneous approximation of a function and its derivatives, and these tools have been applied to the nonlinear modeling of communication devices, such as MESFET or HEMT transistors.

 

Publications

A list of publications can be found at:

• Personal profile at ORCID
• Profile at Web of Science (ResearcherID R-4021-2018)
• Profile at Scopus (Elsevier)
• Personal profile at Google Scholar
• Profile at DBLP computer science bibliography

 

Software

Some software:

• Bayesian neural network for binary classification: MATLAB code (See Readme.txt for details)

2024 Marcelino Lázaro