Statistical Modeling and Inference Group
Information Theoretic Methods Team in SPAG Center of Excellence
The group is involved with developing advanced mathematical and statistical modeling tools in various areas of signal processing. The considered tasks include especially signal compression, signal analysis, model structure selection, hypothesis testing, sparse modeling, and feature selection. The signals we deal with are coming from various areas, most often being audio, image, bioinformatics, and biomedical signals. The group contributes with methodological advances by developing new algorithms and providing new insights into theoretical open problems. It also seeks comprehensive high performance solutions to selected applications, e.g., providing the state of the art results in compression of various signals.
Current research of the group can be mainly divided into the following four topics:
- Information Theoretic Methods
- Signal and Data Compression
- Convex Optimization and Sparse Modeling
- Genomic Signal Processing.
Each of these research topics is described in more detail in the following sections.
Information Theoretic Methods
Information theoretic methods are a well established area, known in the first place for providing solutions to the problem of optimal structure selection of signal models, which is a topic of paramount importance for advanced modeling applications. One main line of research is towards developing the theory of stochastic complexity for linear regression, sum of sinusoids, AR, and ARMA models and also for composite hypothesis testing. The computation of normalized maximum likelihood models for classes of discrete models is another main direction which was pursed with applications in engineering and bioinformatics for sequences and time series data, and which is now extended toward imaging applications.
Signal and Data Compression
Signal and data compression are pursued as a major activity directed towards proposing the most efficient ways to encode various signals (audio, speech, images, bioinformatics sequences, ECG) mainly in a lossless manner, but also in a lossy ways for very low bitrates. The highlights include lossless audio coders and a genome compression program which constitute the state of the art in their field. We also use the compression expertise to propose the most efficient implementable minimum description length representation of signals and data, with applications to analysis and statistical inference.
Convex Optimization and Sparse Modeling
Convex optimization gained significant practical importance at the end of the 90s, with the advent of semidefinite programming libraries. Our expertise in this field was used in the last 10 years especially in the manipulation of positive trigonometric polynomials, with applications like FIR, IIR, filter bank design, MA spectral estimation, and others. Recently, convex optimization has met sparse modeling, providing a tool of choice for solving problems where the solution vector is known to have only few nonzero elements. However, it is in the alternative approach of greedy algorithms that we have found efficient solutions to adaptive filtering problems, with possible applications in lossless audio coding.
Genomic Signal Processing
We were among the initiators of a new discipline, genomic signal processing, which utilizes methodologies from signal processing field for solving problems in bioinformatics, where the volume of data produced by the new exploration technologies requires advanced and efficient modeling techniques, for handling (compression), processing (feature extraction and estimation), and finally for inference, with applications, e.g., to cancer diagnosis and staging.
Members
Projects
The group is involved in cooperation with academic and industrial partners from various areas of engineering, biology, and medicine.








