Tom Kevenaar  (Chief Architect priv-ID Biometrics)

Classification Theory for Biometric Authentication

The field of biometrics is concerned with recognizing individuals by means of unique physiological or behavioral characteristics such as fingerprints, faces or irises. Before an individual can use a biometric system, a biometric is measured, processed and stored somewhere in the biometric system. During authentication, a live measurement is compared with the stored reference infomation and the systems determines if the authentication is successful. All the measurements in a biometric system are inherently noisy, mostly due to varying interaction of the biometric with the biometric sensor. Therefore, comparison of biometric measurements must be treated as a statistical classification process that determines if a measured biometric is drawn from the probability distribution of the claimed identity (the genuine distribution) or from the distribution describing all other individuals (the impostor distribution or background distribution). In this presentation, a general architecture of a biometric system will be given as well as some of the often-used underlying statistical processing methods such as PCA, Fisher LDA, likelihood ratio, Euclidean distance and correlation. We also explain the basic ideas behind Support Vector Machines and the problems that occur with few samples in high dimensional spaces.

Dee Denteneer (Philips Research)

Throughput limitations in CSMA-type networks

abstract: We consider abstract models for CSMA-type packet radio networks in which each radio accesses the medium according to a Carrier Sense Multiple Access (CSMA) protocol, identically and idependently of the other radios. We show that this protocol leads to either extreme unfairness between the radios or poor utilisation of the medium in networking scenarios. From a mathematical perspective, we discuss some new mathematical models related to Dijkstra's classical Philosophers' problem, and state theorems and conjectures concerning these.

Prof. dr. ir. Sem Borst (Lucent and Eindhoven University of Technology)
Channel-aware scheduling and user mobility in wireless data networks

Channel conditions in wireless networks exhibit huge variations across space and time, giving rise to vast random fluctuations in the feasible transmission rates. Channel-aware scheduling strategies provide an effective mechanism for improving throughput performance by exploiting such rate variations, and have been extensively investigated at packet level for a static user configuration. In this talk, we explore the performance implications at flow level for a dynamic user population, taking into consideration rate variations on a slower time scale and wide-range user mobility as well.

First of all, we present simple necessary conditions for flow-level stability, and prove that these are in fact (near-)sufficient for a wide family of utility-based scheduling strategies. It is further shown how the flow-level performance of the Proportional Fair scheduling strategy may be evaluated by means of a multi-class Processor-Sharing model with state-dependent service rate. In addition, we examine the impact of rate variations on a slower time scale, and establish that two limit regimes, termed fluid and quasi-stationary regime, yield explicit, insensitive performance bounds.

Finally, we turn attention to a network of several base stations with hand-offs of active sessions governed by wide-range user mobility. It is demonstrated that mobility increases the capacity region, not only in case of globally optimal scheduling, but also when each of the base stations adheres to a local fair sharing discipline.

dr. Andreas Loepker (EURANDOM)
The asymptotic behavior of the maximum process and the first passage time of Markovian growth collapse models

Abstract: Subject of the seminar is the asymptotic behavior of the running maximum process and the first passage time of Markovian processes with deterministic increase and random downward jumps. Examples of such processes are population growth processes and the window size process in TCP
(Joint work with Wolfgang Stadje, Osnabrueck & Johan van Leeuwaarden)

Pieter de Bokx (MiPlaza Materials Analysis), Gerald Lucassen (Biomedical Photonics, Philips Research)
Multivariate Data Analysis Applied in Molecular Diagnostics

In the Healthcare program at Philips Research, various methods are under investigation for diagnostic tests based on specific protein or DNA detection. A common trend in the field of molecular diagnostics of infectious diseases is the need for rapid detection of multiple analytes. Implementation of these so-called multiplex methods requires chemical analysis techniques that allow simultaneous collection of multiple analyte data, and data analysis techniques that allow unscrambling of the compound signals into those of the individual constituents.

In the presentation we will highlight an optical method that uses surface-enhanced resonance Raman spectroscopy (SERRS) for the analysis of DNA from clinical samples. The SERRS method is suitable for simultaneous measurement of multiple analytes in solution. As multiplex methods involve observation and analysis of more than one statistical variable at a time, multivariate statistical methods are required, both in experimentation and in data analysis. In the talk we will present the experimental design and multivariate data analysis techniques that we have used to reach clinically relevant results from measured SERRS data.

dr. Ashish Pandharipande (Connectivity Systems and Networking group; Philips Research), November 7, 2007
Cognitive Wireless Systems

Traditional models of spectrum management were based on exclusive allocation of
spectrum for designated wireless systems to ensure interference protection from
and to other wireless systems. Such exclusive license allocation however leads
to poor spatio-temporal utilization of spectrum. Cognitive radio is a recent
technology that aims at alleviating this problem. A cognitive radio is capable
of identifying pieces of licensed spectrum that are not in use by a licensed system
at a given time and location, and can access such spectrum dynamically without
causing harmful interference to the licensed system. In this talk, we focus on
these two aspects, namely spectrum sensing and dynamic spectrum access. In the
first part, performance improvements achieved using multiple antennas in a
cognitive radio for sensing will be presented. In the second part, a cognitive
relay system based on spectrum pooling will be presented, and the problem of optimum
spectrum pool reassignment discussed.

Dr. Martin F. McKinney, May 2, 2007
Music Content Analysis: Extracting the style and mood from musical audio

We have developed a set of features, derived from audio waveforms, that allow us to model and distinguish various musical styles and moods. The features, based on signal properties, musicological statistics and perceptual factors, are divided into four groups: spectrotemporal features, tonality features, percussiveness features and rhythmic features. We apply traditional pattern classification techniques (Quadratic Discriminant Analysis) on the feature data to train models for music style and mood classification. Logical extensions to this work would include more sophisticated feature selection and ranking procedures, advanced classification and feature extension approaches, (e.g., support vector machines), as well as new techniques for feature space visualization.

Dr. Radu S. Jasinschi, May 2, 2007
Markers for Neurodegenerative Diseases

We will present methods for the segmentation and representation of brain information based on magnetic resonance imaging data. This data comes from two clinical studies focusing on ageing related neurodegenerative diseases: (i) atherosclerotic risk factors (PROSPER) and (ii) Alzheimer's disease (Memory Clinic). We will discuss markers for neurodegenerative diseases and how they are related to clinical data.

Wim Verhaegh (Philips Research Laboratories), February 21, 2007
Data Mining for Biomarker Discovery -

In the healthcare area, the increasing possibilities to perform molecular measurements on patient samples, enable whole new ways to diagnose patients, and to predict their status and response to treatments. To do so, one has to investigate what the relation is between these molecular measurements and the particular diagnosis or status of patients. A problem in this is that gene expression microarrays typically yield tens of thousands of measurements, and mass spectrometers to measure the presence of proteins give similar numbers of measurements, whereas typical patient studies concern a few hundreds of patients and controls. Furthermore, the measurements often contain quite some noise. This makes the risk of finding spurious patterns significant, and finding relevant features (biomarkers) a challenging task.

In the presentation, I will give a flavor of the data mining problem and how it is approached. To this end, I will first give a short introduction into gene expression microarrays, and next go through the different steps in an analysis on breast cancer as presented in the literature. Next, I show a few preliminary experiments we did, which clearly indicate the risk of overfitting, followed by some issues reported in the literature.
 

Bart Bakker (Philips Research Laboratories), February 21, 2007
In-Silico Models for Clinical Diagnostics -

Current molecular measurement techniques (like microarrays, bioassays, etc) have provided a substantial pool of knowledge about many different processes that take place inside the human body. For many diseases it is known which of these processes (or pathways) are involved, and thus it is theoretically possible to describe an entire disease on a molecular level. Such a description, or -disease model-, could be of great value in clinical diagnosis (disease recognition), prognosis (predicting what is going to happen to a patient, e.g. after treatment) and monitoring (continuously checking whether the patient is still in a stable/safe condition).

Constructing a disease model, however, is easier said than done. Many interacting molecules play a role, not all interactions are known, and those that are known are generally described on a qualitative level only. A sound mathematical and algorithmic framework is needed to translate the multitude of scattered, incomplete, biological models into one clear and robust, numerical computer model. I will sketch the outline of this effort, and focus especially on probabilistic methods. The latter are of special importance for this type of modelling, since the models are built on uncertain a priori knowledge (vague or contradicting publications) and very noisy molecular measurements.