Cognitive Neuroscience (neurocognition)
Modelling attention in the visual system
Introduction
In the last decade we have learned a lot about the visual system. Visual cortex is involved in object recognition and many other visual tasks. To achieve this, there are massively parallel feedfoward networks from low level cortical areas to high level cortical areas. There are also massively parallel feedback networks in the other direction, whose role is not completely clear at the moment. One remarkable hypothesis is that a visual percept is represented in different araes of the brain, position in one area, shape, colour and motion in other areas, etc. Such compositional representations have clear advantages: multi-object visual scenes are represented efficiently and may facilitate learning.
The Binding Problem
Although efficient, such a representation create binding problems. In our work we consider visual attention as a mechanism for solving such binding problems. It creates a dynamic link between different visual attributes of the same object if one of the objects is relevant for behaviour. For example the presence of the colourred may trigger an examination of the nature of the red object: red may indicate "edible" or dangerous. It is important that the red is associated to the right object.
Other aspects of visual attention
Attention is often introducedas a mechanism to selectively improve processing of locations of the visual field (spatial attention) or objects (object-based) attention that are behaviourally relevant. Such considerations are important as well. They are partly complementary to and partly part of the mechanism that solves binding problems. A further discussion is given here.
The architecture and dynamics of visual cortex
In order to test our ideas on visual attention, modelling is extremely important. An important activity of the Biosystems group is to establish biologically plausible models of visual attention: models that are as realistic as possible in the description of the architecture and neuronal dynamics of visual cortex. You can find a more eloborate discussion here. This automatically leads to the following research theme:
Modelling neuronal populations
Neuronal dynamics
Given the increased spatial and temporal resolution of modern imaging techniques, a realistic description of the neuronal circuit becomes important. Artificial Neural Networks may be appropriate for modelling some cognitive phenomena, but in order to describe the dynamics of cognitive processes, it is necessary to model the activity of neuronal populations at an appropriate level. For cognitive models, the individual neuron is not the appropriate level. Interestingly, the last decade has seen a strong development of so-called population density techniques, which describe groups of spiking neurons. We expect that this mesoscopic level of description will be more appropriate to relate models to imaging data. Within the group some highly efficient and stable algorithms have been developed for solving the Fokker-Planck equations which typically result from population density approaches.
The MIIND project.