Schedule ISLA Colloquium 2006
Every talk runs from 11.30 till 12.30 and is held in
room F0.13 Kruislaan 403.
If you click on the titles in the table below you will see the abstract of the talk.
More information is found on the schedule page of the ISLA SOOS.
Speakers, their titles and their abstracts
- Jan 17
- Jan-Mark Geusebroek
- Jan 24
- Balder ten Cate
- How to design nice languages
- Many issues play a role in the design of logical languages (such
as query languages, knowledge representation formalisms, and
specification languages). There is a well known trade-off between
expressivity (how much you can say in a language) and complexity
(how difficult it is to reason about expressions in the language),
but there are also many other, less known, mathematically
interesting interactions between desirable properties of such
Taking the XML path language XPath as an example, I will discuss
some of these properties and the way they interact with eachother.
In doing so, I will cover some recent work of Maarten Marx and
myself, as well as plans for my VENI project, which has just been
- Feb 7
- Theo Gevers
- Color in Computer Vision
Color information has become a very useful information cue in computer
vision. In this talk, an overview will be given on the use of color in the
area of feature extraction, object recognition, and visual tracking.
Physical models and learning techniques are discussed to obtain invariance
and discriminative power. Further, the detection and classification of
local structures (i.e. edges, corners, and T-Junctions) in color images
is discussed. Finally, a few applications are presented in the context of
image/video retrieval and player tracking in sport video.
- Feb 28
- Nikos Vlassis
- Bayesian Reinforcement Learning
Reinforcement Learning (RL) refers to a class of methods addressing the
problem of agent decision making under uncertainty. This problem can
often be formalized as a Markov decision process (MDP), in which case
optimal decision making corresponds to computing an optimal policy of
this MDP. When the environment model (stochastic dynamics) is known, an
MDP optimal policy can be easily computed by dynamic programming. When
the environment model is unknown, the agent must learn its policy online
by 'exploring' the environment. But what is the best way to explore an
In this talk I'll describe Bayesian RL, an approach that in principle
allows us solve the RL exploration problem optimally. The main idea is
to form beliefs over the space of all possible environment models, and
compute an optimal policy over that belief space. However, searching for
policies over an infinitely dimensional belief space of models is a
daunting task, making Bayesian RL a notoriously difficult problem to
solve. In this talk I'll present some recent developments in the theory
of Bayesian RL, which open new possibilities for the development of
- Mar 21
- Arnold Smeulders
- Objects and objections
In the talk I discuss the notion of an object in computer vision. The
segmentation of a scene is described as the delineation of objects in the
image field and generally believed to be the first processing step of
human and computer vision alike defining the presence or absence of
objects in an image. This is in contrast to the current wide believe in
computer vision that universal algorithm for segmentation do not exist or
are very hard anyway. In this ongoing debate I give a definition of
segmentation and an overview of different brands of segmentation
- May 23
- Marcel Worring
- Browsing Visual Collections
- The semantic gap between low level visual data and conceptual
interpretation of the same data dictates that interactive browsing is an
essential element of any visual information system. Such interactive
browsing is a complex interplay between analysis of the data,
visualization and user feedback. To structure the design of such methods
we have developed the query space in our "CBIR at the end of the early
years" paper. This query space is composed of the active dataset,
features, a similarity function, interpretations and the relations among
the four components. Since the early years several new insights in the
characteristiscs of the query space have been gained. In this talk we will
elaborate on the relation between different tasks, the role of large scale
high-level indexing and new feedback and visualization methods.
- May 30
- Maarten de Rijke
- Consumer Generated Text
- Recent years have seen a rapid increase in the amount of consumer
generated content that's available on the web. YouTube (for
video sharing) has now entered the top 25 of most visited web
sites. Before that Flickr rose to fame with its photo sharing
facilities. And before that blogs had arrived on the scene. The
massive availability of consumer generated content offers many
challenges and opportunities for anyone interested in developing
search, discovery, and analysis tools for web content.
In this talk I will deal with consumer generated text, and, more
specifically, with blogs. I will report on two research lines
that are being pursued within the ILPS group. One deals with
moodviews.com, a collection of tools for tracking streams of
mood-annotated blog texts. The other deals with hierarchical
language models for capturing a blogger's profile. One application
of these models is in contextual advertisement on personal blogs:
state-of-art ad placement methods for non-blog web content are
not effective for personal blogs. An alternative method based on
hierarchical language models proves to be much more effective.
- June 13
- Ben Kröse
- Digital Living: challenges for computational intelligence research
There is an increase of ICT in residential environments: smart homes
are becoming popular, in particular for monitoring purposes in
elderly care. The systems which are being used consist of many sensors
distributed in the home, and methods from the field of computational
intelligence are used for data analysis.
In the presentation I will focus on three aspects of these smart homes.
First I will focus on algorithms for behaviour analysis of the humans
and the relation with earlier work on multi camera surveillance.
Then I will report on the Cogniron project, where we study (cognitive)
representations for the interaction with humans. Thirdly I will present
results of a study on the acceptance of ICT (the Philips 'iCat') by
- July 4
- Cor Veenman
- Forensic Intelligence: Pattern Recognition in Forensic Practice
In the forensic practice, the number and size of available databases is
growing rapidly. These databases range from biometrics of convicted
people, to case report databases and case related environmental or drug
samples. Forensic Intelligence deals with processing these databases in
order to obtain useful hypotheses and to evaluate available evidence.
In this talk I will give some research directions and examples of work in
progress. First, multi-modal biometrics is an important direction for
further research. Especially in forensic applications partial profiles are
common and combined biometrics enable to narrow down the group of possible
suspects. Another type of research is the prediction of drug types based
on its composition in elements. This results in high dimensional
classification problems that require specialized models.
- September 26
- Sennay Ghebreab
- Brain reading: decoding mental states of humans from functional
- In cognitive neuroscience, a trend can be observed toward psychophysical
experiments involving naturalistic scene stimuli rather than simple
controlled stimuli such as images of lines with predefined orientations.
In computer science, a trend can be observed toward systems that enable
automatic scene characterization and interpretation on the basis of
objectively defined (statistical) scene features and classification
schemes. In neuroimaging analysis, a trend can be seen toward
pattern-based data analysis taking into account the full spatial pattern
of brain activity rather than concentrating on specific brain regions.
These trends are rapidly converging to a new research area: decoding
mental states of humans from functional neuroimaging data. This research
area aims at identifying, understanding and simulating processes and
representations involved in the interpretation of natural stimuli such as
real world scene by combining psychophysical and computational experiments
with multivariate pattern analysis techniques. In this talk I will
elaborate on our research in mapping brain activation data to natural
sensory stimulation data and report on our participation in the brain
- November 7
- Dariu M. Gavrila, DaimlerChrysler R&D, Ulm, Germany
University of Amsterdam, Amsterdam, The Netherlands
- SMART CARS (THAT SEE/SAVE PEDESTRIANS)
I'll start with a short overview of the rapid developments
in the intelligent vehicle field over the past 2-3 years.
The main part of the talk then covers recent progress at
DaimlerChrysler on a viable vision-based pedestrian
safety system . I will focus on the pedestrian classification
sub-problem, describing a recent study  which examined
multiple feature-classifier combinations with respect to
their ROC performance and efficiency. The underlying
dataset, involving several thousands of pedestrian
samples is downloadable as part of a new-to-be-established
benchmark on pedestrian/object classification. I furthermore
discuss semi- and fully automated techniques for enlarging
the training set w.r.t. the target class.
The performance of the DaimlerChrysler pedestrian system
was extensively tested in real urban traffic. But I also
present the outcome of some remarkable PreCrash experiments
on the test track (with real pedestrians) ...
 D. M. Gavrila and S. Munder. "Multi-cue Pedestrian
Detection and Tracking from a Moving Vehicle". To
appear in IJCV, Feb/March 2007.
 S. Munder and D. M. Gavrila. "An Experimental Study
on Pedestrian Classification", PAMI, vol.28, nr.11, 2006.
- Tuesday 21 November 2006
- Arnold Smeulders, Max Viergever, Marcel Reinders and others
- Fall Meeting "25th Anniversary of the NVPHBV"
*Fall Meeting "25th Anniversary of the NVPHBV" *
Date: Tuesday 21 November 2006
Local organizer: Informatics Institute of the University of Amsterdam
Address:, CWI - lecture room Z.011 Kruislaan 413, Amsterdam
We cordially invite you to attend the fall meeting of the NVPHBV in
celebration of its 25th anniversary. The program starts at 10.00 and
includes three invited keynote presentations, several oral presentations
by NVPHBV members centered around the 25th ledenvergadering. We start
the day with coffee (09.30), serve lunch and end with drinks. We will
also distribute the CDROM "Fifth Quinquinnial Review" to which many of
you have contributed.
Invited keynote speakers
Professor Arnold Smeulders (University of Amsterdam)
Professor Max Viergever (University Medical Center Utrecht)
Professor Marcel Reinders (Delft University of Technology)
Call for papers
Besides the keynote lectures we have several slots for presentations by
our members. Please submit a title (with abstract) to
email@example.com (with a cc to firstname.lastname@example.org) before 1
November 2006. The definitive program will be sent to you on 6 November
Please register as soon as possible, but no later than 17 November 2006,
by sending an email to Mrs. Virginie Mes
- Leo Dorst
- Frans Groen
- Intelligent Autonomous Systems:
Past, present and future?
The well-known sense-think-act loop is characteristic for autonomous systems. Closing this loop enables analysis of the result of the action and so learning of goal directed behaviour. To create an autonomous system means that two essential components are needed besides intelligence: perception and actuation. Autonomous systems can be biological inspired, that forms a proof of existence, but on the other hand an artificial technical system gives also the freedom to incorporate non-human sensing.
In the past the focus has been on single agent systems, which learn certain tasks. The development towards real-world multi-agents systems was reflected in our group by the participation in RoboCup and in the DECIS lab. Multi-agent collaborative systems results in distributed sensing and requires distributed decision making. When no direct decision making is involved the first aspect is closely related to sensor networks. The complexity of the problem increases exponentially with the number of agents, so how to come up with approximate methods to solve this problem is an essential question.
You could debate whether autonomous systems are what you need in real-world applications. In the end the human is in control and systems are only in certain aspects autonomous (e.g. in sensor management and sensor data fusion). The trend is towards the cooperation of human and artificial agents. This cooperation can be present both in sensing and in the assistance to human decision making.