DTAI

Monday December 3 2012 at 16h30 in Celestijnenlaan 200A (auditorium 00.225)

The Relax, Compensate and Recover Framework for Approximate Inference
by Guy Van den Broeck (PhD student DTAI)

Relax, compensate and recover (RCR) is a framework for approximate reasoning that has succesfully been applied to many different inference tasks. It is based on three simple concepts. First, it simplifies the problem at hand by relaxing equivalence constraints between copies of variables in the model. Second, it compensates for this loss of information by reparametrizing the model. Third, it incrementally recovers relaxed equivalences that significantly improve approximation quality. The RCR framework has been applied to (weighted) MAX-SAT problems, several Bayesian network inference and learning tasks and to lifted inference in statistical relational models, leading to state-of-the-art algorithms in each case. Furthermore, it provides a unifying semantics for many existing approximation algorithms.

Friday 30 November 2012 at 14h00 in "Auditorium Kasteel" in the Arenbergkasteel, PhD defence by Theofrastos Mantadelis:

Efficient Algorithms for Prolog Based Probabilistic Logic Programming

The integration of probabilistic reasoning with logic programming has become one of the challenges in Artificial Intelligence. Efficient probabilistic logic programming is critical for applications like: mining biological databases, classifying web pages, simulating network protocols. Lately, a lot of Probabilistic Logic Programming (PLP) formalisms have surfaced. Given that PLP is the combination of logic programming and probabilities which are two very different fields, it is expected that researchers from several fields come up with different approaches to tackle the presented challenges. This has resulted in a new discipline called Probabilistic Logic Learning (PLL) or Statistical Relational Learning (SRL) and a very active research community.

Within this community, ProbLog a probabilistic extension of Prolog, has appeared. ProbLog was motivated by the task of mining links in large probabilistic graphs. The simple but powerful ProbLog formulation was extended in order to support inference on several different models. Soon ProbLog evolved into a general purpose probabilistic programming language that provides infrastructure for many PLL/SRL tasks. The two most critical aspects of a PLP language are its expression power, and its scalability over common PLL problems.

Monday November 26, 2012 at 16h30 in Celestijnenlaan 200A (auditorium 00.225)

Learning from Geometrical Data
by Thomas Fannes (PhD student DTAI)

The class of geometrical data is an interesting class as one encounters them in real world applications, e.g., the representation of images, geographic information systems, traffic networks, etc.
Despite the interest in this type of data, there has been no in-depth study of the theory of learning related to geometrical data. In this presentation, we will look closer to two subtasks in this area: Firstly, we present a number of learnability results for plane and planar graphs and we list a number of important prediction applications in the context of applications such as image recognition and spatial reasoning. Secondly, we take first steps towards analyzing the learnability tasks when considering projections of solid objects. In particular we consider the learning tasks where the input is a set of random generated orthogonal projections of polygons. We examine unique reconstructability of the projected model and present probably approximately correct techniques to learn the model. Finally we outline the future research needed to acquire a good understanding of these tasks, as well as possible solution methods.

Monday November 12 2012 at 16h30 in Celestijnenlaan 200A (auditorium 00.225)

Discovering Patterns in Sequences
by Nikolaj Tatti (postdoc DTAI)

Discovering patterns from sequential data is an active and well-studied subfield of pattern mining. Sequential patterns are, in essence, sets of events that co-occur often enough in their vicinity.  In this talk we will go through basic definitions involved with sequential patterns. Namely, we will discuss different types of data and different types of sequential patterns, how such patterns should be scored and discovered efficiently. We will also discuss pitfalls related specifically to discovering sequential patterns. Finally, we will discuss on how to apply statistical methods in order to assess statistical significance of discovered patterns.