Shared beliefs contain information which is part of the common ground for the group [3]. The beliefs can then be directly exploited by high-level cognitive functions such as planning, cross-modal learning or communication. Attributed beliefs are beliefs which are ascribed to other agents. The temporal dimension can be adapted in a similar way. To be able to interact naturally with humans, robots needs to be aware of their own environment. We leave this question as an interesting topic for future research. Beliefs are thus organised in a complex two-dimensional structure, with horizontal relations between belief dependents and vertical relations between belief relatives.

Key to our approach is the use of a first-order probabilistic language, Markov Logic [24], as a uni- fied representation formalism to construct rich, multi-modal models of context. For a pair of two percepts p1 and p2 , we infer the likelihood of these two percepts being generated from the same underlying entity in the real-world. In Perception and Interactive Technologies: The outcome is a set of possible unions, each of which has an existence probability describing the likelihood of the grouping. In the case of world x, the first formula is violated, while the second is not. Bottom-up belief model formation.

Learning words and syntax for a scene description task. The research in his PhD was mainly dedicated to Bayesian spatio-temporal modeling of bycatch in the Barents Sea shrimp fishery, and was conducted in collaboration with the Norwegian Marine Research Institute.

This perceptual grouping process is triggered at each insertion or update of percepts on the binder provided the number of modalities in the Levels of beliefs Stable belief Temporal union Temporal smoothing Multi-modal belief Percept union Percept Tracking The similarity of a pair of beliefs is based on the correlation of their content and spatial frameplus other parameters such as the time distance between beliefs.


pierre lison thesis

The Markov Logic Network for tracking works as follows. I am now back in Oslo after a 3-months research stay at the Idiap institute in Switzerland. The core of the binder is its working memory, where beliefs are formed from incoming per- ceptual inputs, and are then iteratively fused, refined and abstracted to yield stable, high-level beliefs.

The construction of such belief models raises two important issues for the system developer. Our approach departs from previous work such as [13] or [27] by introducing a much richer modelling of multi- modal beliefs. The main purpose of this project was to predict future bycatch for performing cost effective regulations.

Following the fusion operation, beliefs are then gradually refined — new, improved esti- mations are derived for each belief feature, given the collection of knowledge sources which have been merged.

pierre lison thesis

Bottom-up belief model formation. They can also be used by perceptual components to adapt their internal processing operations to the current situated context contextual priming, anticipation, etc.

In the simplest case, the spatial dimension can be modelled by a discrete set of regions and the temporal dimension via intervals defined on real-valued time points. Information fusion for visual reference resolution in dynamic situated dialogue.

Representation, reasoning, and relational structures: The binder is directly connected to the other subarchi- tectures i. SAFERS investigates how speech technology and machine learning can be used to ljson transcribe emergency calls in Norway and provide real-time analysis and prediction tools to emergency response services.

pierre lison thesis

In gen- eral, each alternative value can be expressed as a propositional logical for- mula. Given the requirements of our application domain see Section 1and particularly the need to operate under soft real-time constraints, such approximation methods are an absolute necessity.


New Research Scientists have joined SAMBA | Norsk Regnesentral

This can be done by breaking down the formulae into ele- mentary predications, and assuming conditional independence between these elementary predicates. Generation and evaluation of user tailored responses in multimodal dialogue.

A new approach to object-oriented middleware. It is important to note that beliefs can express past or future knowledge i. Section 2 provides a brief intro- duction to Markov Logic, the framework used tuesis belief modelling. Belief modelling for situation awareness in human-robot interaction Attributed beliefs are beliefs which are ascribed to other agents.

Pierre Lison Completes Doctoral Degree

Finally, subarchitectures can thesix communicate with each other by accessing reading, inserting, updating their respective working memories. An edge between two nodes signifies that the corresponding ground atoms appear together in at least one grounding of one formula in L.

The epistemic status of this information is attributed. Huynh and Raymond J.

Goal-driven components can be controlled explicitly at runtime by a task manager specifying which component is allowed to run at a given time. The thwsis probability distribution of the Markov network can then be factorised over the cliques of G: Several machine learning algorithms for parameter learning thfsis be applied to this end, from classical gradient-based techniques to more sophisticated algorithms specifi- cally designed for statistical relational learning [21,12].

Such algorithms are able to fix an upper bound on the amount of computation required for any processing operation.

Author: admin