Intelligent Systems Lab Amsterdam


Please see the new ISLA website at: http://isla.science.uva.nl

ISLA AIO Soos

See the schedule page for more information on the AIO soos.

Schedule ISLA Colloquium

Every talk runs from 11.30 till 12.30 unless otherwise stated and is held in room F0.13 Science Park 107.

If you click on the titles in the table below you will see the abstract of the talk.

Date Speaker Title
July 13, 2010 Sarvnaz Karimi Search in Biomedical Literature
June 8, 2010 Zoran Zivkovic How smart is my new LCD TV?
May 11, 2010 Sonja Smets From dynamic belief revision to dynamic rationality Special Location: F3.20
April 13, 2010 Robby Tan Introduction to Physics-based Computer Vision
March 16, 2010 Jaap van den Herik We May Solve Chess: What about Go?
February 16, 2010 Frank Nack & Vanessa Evers Mobility as a Challenge for Interactive storytelling AND Social Intelligence: Informing the next frontier of Artificial Intelligence
January 19, 2010 Siti Norul Huda Sheikh Abdullah Determining Adaptive Threshold Value for Image Segmentation
November 17, 2009 Arnold Smeulders The Impact of ICT
October 20, 2009 Marco Wiering Reinforcement Learning for Classification Problems
September 22, 2009, (10:30-11:30) Bert Kappen Graphical models and approximate inference: from medical diagnosis to stochastic control theory.
April 28, 2009 Jacob Verbeek Discriminative Metric Learning in Nearest Neighbor Models for Image Auto-Annotation
March 31, 2009 Jun Wu Online Learning in Video Concept Detection
March 31, 2009 Ben Krose From Bayesian Theory to a Business with Camera Systems
March 3, 2009 Peter Hall A face is face is a face
February 3, 2009 Daniel Fontijne Towards markerless motion capture using geometric algebra

Speakers, their titles and their abstracts

Date
July 13, 2010
Speaker
Sarvnaz Karimi
Title
Search in Biomedical Literature
Abstract
Complicated Boolean retrieval is still the dominant search paradigm in the clinical and biomedical domains. Nurses, medical students and medical researchers are trained to formulate Boolean queries for their daily search in biomedical texts, and if they fail to find any relevant information, health librarians may come to their rescue. Sometimes, however, finding any relevant evidence is difficult, even for expert users of the existing systems (such as PubMed or Ovid MEDLINE). In this talk, I'll review some of the research results from our group on improving search in biomedical literature. Our research outcomes address two main problems that have made it challenging for any alternative approach to complicated Boolean retrieval to be practical: (1) focused information needs which are hard to formulate, because a query should cover multiple criteria that need to be satisfied for a relevant research article to be retrieved; and (2) the strict requirement for finding all the existing evidence. I'll also introduce unsolved problems and possible approaches that we are investigating, including finding the diversity of the Boolean queries output for better query formulation and workload reduction.
Date
June 8, 2010
Speaker
Zoran Zivkovic
Title
How smart is my new LCD TV?
Abstract
High contrast images on a large LCD screen can look very impressive. However, a careful observer might soon notice that the picture quality becomes less pleasing when moving objects are displayed. Motion might not appear smooth and the moving objects might look blurred. Certain properties of the LCD display make human visual system much more sensitive to such artifacts. This talk will give an overview of the complex processing introduced in order to reduce these artifacts inside high-end TV sets such as: motion estimation, occlusion detection, spatial/temporal super-resolution, film material detection.
Date
May 11, 2010
Speaker
Sonja Smets
Title
From dynamic belief revision to dynamic rationality
Abstract
My talk (based on joint work with Alexandru Baltag and Jonathan Zvesper) is about using recent developments in Logic to better understand and model "rationality" in extensive games, by taking into account the dynamics of belief. The main idea is that, in order to correctly factor in the evolution of players' beliefs (about each other) throughout the game, we need a novel notion of "dynamic rationality". This is a context-dependent (time-dependent), knowledge-dependent, belief-based and "future-oriented" concept, that presupposes what one may call "epistemic freedom of choice". To formalize this notion, I use a belief-revision-friendly version of Dynamic Epistemic Logic. I then apply this concept of dynamic rationality to propose a new solution to a famous debate (between Aumann on one hand, and Stalnaker and Reny), concerning the epistemic conditions for the so-called backward-induction solution. "Backward induction" is the oldest, simplest and perhaps the most natural solution concept in Game Theory. But the reasoning underlying this solution seems to give rise to a fundamental paradox (the so-called BI paradox). I use the concepts developed in this talk to address the paradox, and I argue that the correct epistemic condition underlying the backward induction method is more general and weaker than Aumann's: "common knowledge of stable belief in (dynamic) rationality".
Date
April 13, 2010
Speaker
Robby Tan
Title
Introduction to Physics-based Computer Vision
Abstract
The world contains abundant visual information. The main task of computer vision is to acquire, extract and interpret that information into meaningful representations. Over decades of development, computer vision has become a field with multiple aspects and approaches emerging from diverse disciplines and applications. One of the important approaches is physics-based computer vision (or physics-based vision, for short) which in principal utilizes physical models of the world to extract visual information. This approach has become one of the core topics in computer vision, since visual information is in fact spatial collections of light rays that are emitted, transmitted, scattered, absorbed, and reflected according to physical laws. While it has significantly contributed to the development of computer vision, physics-based vision has also played crucial roles in computer graphics and virtual reality, since to arrive at realistic image rendering, one should employ physical models of the world and their parameters.
Date
March 16, 2010
Speaker
Jaap van den Herik
Title
We May Solve Chess: What about Go?
Abstract
Currently, we are in the fourth phase of computer chess. The phases are: (1) playing an acceptable game, (2) playing a game at grandmaster level, (3) playing above World Champion level, and (4) solving the game. The prevailing question is: can a computer solve the game of chess? The number of different reachable positions is 1046 (Chinchalkar, 1996). However, arranging the list of positions in such a way that they can be visited or cut off by a clever computer program requires (1) a new intelligent approach, (2) a considerable speed-up of computer power, and (3) a considerable enlargement of storage capabilities. Stimulating developments in this respect are (a) Solving Checkers by Schaeffer et al. (2007), and (b) defeating Kim Myungwan on Go (in a 9-stone handicap match, 2008). New techniques potentially usable in chess are: (i) Monte-Carlo Tree Search, (ii) UCT (Upper Confidence bounds applied to Trees), (iii) Supercomputers, such as IBMp6 (with 3328 processors), and (iv) Grid Technology. Conditions on and predictions for these techniques will be discussed in the lecture as well as their impact. An optimistic date for solving chess is 2035 and a pessimistic one (assuming that we can solve the game) is 2065. The solving time in 2035 (optimistic prediction) will be between 37 days and 4 months. Moreover, for Go (19x19), I predict that the human world champion will be defeated around 2020 by a computer program.

Chinchalkar, S. (1996). An Upper Bound for the Number of Reachable Positions. ICCA Journal, Vol. 19, No. 3, pp. 181-183.

Schaeffer, J. et al. (2007). Checkers is Solved. Science, Vol. 317, No. 5844, pp. 1518-1522.

Date
February 16, 2010
Speaker
Frank Nack & Vanessa Evers
Title
Mobility as a Challenge for Interactive storytelling AND Social Intelligence: Informing the next frontier of Artificial Intelligence
Abstract
Until recently we could access a locations memory mainly through media surrogate s, such as books, drawings, film or audio files, or through face-to-face encounters with people who were able to knit us into the rich but hidden experience fabric of a place. The integration of low-cost pervasive and personal technology in the form of mobile devices and augmented reality into our everyday life starts to change our ex pectations about how to perceive the world around us. We are now able to leave trace s of our emotional or intellectual experience as virtual attachments to any location . As a result we expect that any place, indoors or outdoors, reveals itself to us by confronting us with connection, context, and uncommon perspectives. This presentati on will be an explorative journey into the challenges of such mobile and interactive environments with respect to already existing interactive storytelling technology.

Robots have explored ocean depths, mapped subterranean mines, rescued disaster vict ims, assisted surgeons with operations driven autonoumously across deserts and even explored Mars. The next challenge is society at large. This talk will introduce soci al robotics and a social-psychology experimental approach to inform the design of so cial robot behaviour. The findings of a recent study in effects of robot communicati on style on human robot collaboration will be reported.

Date
January 19, 2010
Speaker
Siti Norul Huda Sheikh Abdullah
Title
Determining Adaptive Threshold Value for Image Segmentation
Abstract
Choosing the wrong threshold can cause failure in either the detection or segmentation process, as the threshold image could either totally hide the required objects or cause worse distraction on the region of interest due to illumination. Heuristic threshold in introduced to find the optimum threshold value. The technique relies on the peak value from the graph of the number object versus specific range of threshold values. A heuristic rule is developed to determine the determine selection of thresholds which are tailored to gray scale distribution of an image. That approach has actually increased the object detection merely up to 99%. However, this approach provides a serial of of selection values, also includes junk objects and it is time consuming when applying onto real time system. Therefore, further improvement on this method is in progress to accommodate real time system specification.
Date
November 17, 2009
Speaker
Arnold Smeulders
Title
The Impact of ICT
Abstract
Rather than talking about some observations on SIFT-filters - which I would like to safe for another occasion - I rather speak on the impact of ICT, previously read at the 15th anniversary of the CTIT in Twente. When asking people how ICT has affected their work, they tend to answer that everything is faster, cheaper, more expensive, or slower. In other words, they tend to answer the question on the processes underlying their profession. This is the automation side of ICT, decisive to many professions and professionals. It has delivered digital prosperity. But I would like to take the question one level further and investigate with you the question: What is the tangible impact of ICT on substance? Can we see and feel the impact of ICT on the earth, on living, on history? Will architects build different buildings? We have to live with impressions, small facts and strictly personal interviews on the impact of ICT.
Date
October 20, 2009
Speaker
Marco Wiering
Title
Reinforcement Learning for Classification Problems.
Abstract
Machine learning algorithms allow a system to learn its knowledge and skills automatically from data or interaction with an environment. There exist three main classes of machine learning algorithms, namely (1) Supervised learning, in which the system receives inputs and the target output values as training examples. (2) Unsupervised learning, in which the system only receives inputs and tries to find important features in the data. (3) Reinforcement learning, in which the system receives an evaluation signal that denotes how well its performed actions are. Although these algorithms are quite different, we show in this talk that it is possible to use reinforcement learning for solving classification problems for which until now only supervised learning algorithms have been used. Furthermore, we show that in our framework, an agent can act to create useful mental representations of the given inputs that enable it to improve the classification performance. We show some preliminary experiments on some well known machine learning datasets.
Date
September 22, 2009
Speaker
Bert Kappen
Title
Graphical models and approximate inference: from medical diagnosis to stochastic control theory.
Abstract
During the last few years, the use of probabilistic methods in artificial intelligence and machine learning has gained enormous popularity. In particular, probabilistic graphical models have become the preferred method for knowledge representation and reasoning. The drawback of the probabilistic approach is that the method is intractable. This means that the typical computation scales exponentially with the problem size, which prevents large scale applications. A popular approximation scheme is provided by belief propagation. This method is closely related to the so-called Bethe approximation from statistical physics.In this talk I will show how this method can be applied to probabilistic reasoning, in particular to two applications: stochastic control theory and expert systems for medical diagnosis.
Date
April 28, 2009
Speaker
Jacob Verbeek
Title
Discriminative Metric Learning in Nearest Neighbor Models for Image Auto-Annotation
Abstract
Image auto-annotation is an important open problem in computer vision. We propose discriminatively trained models for this task. Image tags of test images are predicted using a weighted nearest-neighbour model to exploit labelled training images. Neighbour weights are determined on the basis of the neighbour rank or distance. Our model allows the integration of metric learning to optimally combine a collection of image similarity metrics by directly maximising the log-likelihood of the tag predictions in the training set. In this manner we can automatically find a suitable image similarity that can combine different aspects of image content, such as local shape descriptors, or global color histograms. We also introduce a word specific sigmoidal modulation of the weighted neighbour tag predictions to boost the recall of rare words. We present experimental results for three challenging data sets.We investigate the performance of different variants of our model and compare to existing work. On all three data sets, our models make a marked improvement as compared to the current state-of-the-art.
Date
March 31, 2009
Speaker
Jun Wu
Title
Online Learning in Video Concept Detection
Abstract
In semantic concept detection for online video streams, the underlying data distribution for a certain semantic concept in the visual feature space generally evolves over time. This talk will mainly discuss two key issues: 1) what are the rules underlying the evolving data distribution for different semantic concepts in different conditions? 2) how to update the concept models starting from the limited training samples from the current video sequence? Based on Finite Mixture Models (FMMs), a couple of tracking measures are proposed to describe statistical properties of the evolving underlying data distribution in a quantitative way. According to reasonable assumptions on prior knowledge of the online data streams computed based on these measures, two online learning algorithms are investigated: i) the Multi-granularity Adaptive (MGA) online learning algorithm in supervised learning, and ii) the Online-optimized Incremental Learning (OOIL) algorithm in semi-supervised learning. The experiments show that, the proposed FMM-based tracking measures are useful and they can effectively derive the evolving rule of the target concept, and provide reasonable reference information for the above two online learning algorithms. Furthermore, experimental results on sports video and TRECVID data collections demonstrate that, compared with the existing strategies, these two online learning algorithms (MGA and OOIL) are more effective.
Date
March 31, 2009
Speaker
Ben Krose
Title
From Bayesian Theory to a Business with Camera Systems
Abstract
The IAS group has a lot of projects on the perception and tracking of humans. In this colloquium I will present a particular project that started as a PhD research project. It concerns Bayesian networks for tracking people with distributed camera's. The methods that came out of this were patented. A company that is interested in developing a product on the basis of these metods has a license on the patent. The collaboration with the university has reached a state where an operational system is piloted. Still: reality is more difficult then a scientist can think of...
Date
March 3, 2009
Speaker
Peter Hall
Title
A face is face is a face
Abstract
A face is a "face" whether drawn, painted, photographed or real; that is people can identify object classes under a wide variety of conditions. These conditions are much too wide for state-of-the art computer vision; where the literature is well developed for object recognition, but only from photographic-like images. Little, almost zero, effort has been expended in recognizing objects over a wide gamut of image classes.

The problem addressed in the seminar is: how can a computer learn visual object classes across depiction styles? Our solution is to characterise object structure by a graph of nodes and arc. The seminar explains how to generate a graph from an input image (in any depiction), and use it to both locate object classes and generate novel synthetic art.
Date
February 3, 2009
Speaker
Daniel Fontijne
Title
Towards markerless motion capture using geometric algebra
Abstract
In this talk I will introduce the DASIS project and discuss the low level computer vision work completed in the project so far.

DASIS is an NWO project run by Leo Dorst, Carsten Cibura and myself. The acronym stands for Detection of Articulated Structures in Image Sequences. This more or less means that we want measure the articulated motion of humans and other 'kinematic structures' using regular video images, without a predetermined (human) body model.

The geometry in the project is mostly handled using geometric algebra. In fact, the idea behind the project is to develop geometric algebra in the direction of uncertain, noisy geometric data processing by working on a real-world problem.

Because we aim to use motion as the primary cue for the system, over the past months we have developed software to accurately track features in 3-D using multiple calibrated cameras. These feature tracks serve as input for the next level of the system which we are currently developing. I will discuss the feature tracking software and demonstrate it in real-time. I will also shortly demonstrate the rigid body tracking component which is currently under development to give an idea where the project is going.
Maintained by Vivek E P. Last edited on