|
Workshop Report: S+SSPR 2008 |
|
S+SSPR is a three-day workshop held prior to the International Conference on Pattern Recognition (ICPR 2008, Tampa). This workshop was a collaboration of two IAPR Technical Committees, TC1 on Statistical Pattern Recognition and TC2 Structural and Syntactical Pattern Recognition, and it was organized by Niels da Vitoria Lobo, Takis Kasparis, Fabio Roli, James T. Kwok, Michael Georgiopoulos, Georgios C. Anagnostopoulos, and Marco Loog (Eds). The workshop is a joint workshop that tries to tighten the links between the structural and statistical fields in pattern recognition (PR) research area. The collaboration between structural and statistical PR was reflected in the setup of the presentations: two parallel tracks, one on structural and one on statistical pattern recognition. By the careful synchronization of the starting and ending times of the presentations, it was easy for the participants to jump between either side of field. Around 100 people traveled to the University of Central Florida in Orlando. From the 175 paper submissions 98 papers were accepted: resulting in 56 oral presentations, and 42 posters. On the extensive campus, several presentations, meetings, poster sessions and social evens were organized. Often the events were in different buildings, so the participants had to stroll through the campus from one location to the other. This was especially pleasant, because we had perfect 24C sunny days with blue sky, and the paths were narrow, winding, and it was always a surprise in which building we ended up. The choices for the keynote speakers also reflected this combination of structural and statistical pattern recognition. Two keynote speakers tried to be explicitly on the boundary between structural and statistical pattern recognition. The first keynote speaker was Horst Bunke who discussed graph representations of objects, and how objects can be classifier based on their graph representation. The key ingredient is to compute a dissimilarity matrix between the objects, and to use this dissimilarity matrix for finding a well-generalizing classified. Extra challenges appear when the dissimilarity matrix is not used as 'just' features that characterize an object, but then the true distance-character is exploited. It appears that non-metric feature spaces appear, which introduces all new, interesting problems. The second keynote speaker that tried to mix structural and statistical information into one model was Pedro Domingos. He discussed the topic of 'Markov Logic', which is a combination of (first order) logic and probability. It allows certain logic statements to be expressed in a Markov Random Field. Each logic statement has a weight, or strength, that expresses how strongly this statement has to be satisfied. A Markov Logic Network contains a set of atomic formulas as nodes (each with their weight), and a set of edges that represent functions of the atomic formulas. Using this network, a distribution of possible worlds can be inferred, even when contradictory statements are supplied. In the other oral and poster presentations, the complete range of statistical and structural pattern recognition was present. Talks were given on cluster validity measures, data complexity analysis, multi-level thresholding, soft feature selection, histogram matching, combining similarity measures, combining online classifiers, graph characterization, 3D object representation, etc. To discuss all the topics takes too much time, but one new and intriguing topic was introduced to us by Fabio Roli: Adversarial Pattern Recognition. In adversarial classification, it is assumed that there is an adversary present that actively tries to deceive the classifier. This can happen, for instance, in spam detection, surveillance, fraud detection, and others. In this problem, not only does a good classifier have to be created, but the classifier performance should also be robust against the adversary’s actions to defeat the classifier. A possible approach to confuse the adversary is to introduce a stochastic element to the classifier, such that the classifier randomly switches between two operating conditions. This also forces the adversary to a suboptimal reaction. The wide range of topics in the workshop shows that pattern recognition is very broad and that it has a wide field of application in real world problems. |
|
Tin Kam Ho, winner of the Pierre Devijver Award presenting the Pierre Devijver lecture at S+SSPR 2008
The lecture was titled “Data Complexity Analysis: Linkage between Context and Solution in Classification”
|
|
Proceedings of the conference have been published by Springer-Verlag in Lecture Notes in Computer Science Series (volume number 5342).
Click on the image to go to the publisher’s web site for this volume. |
