[Comp-neuro] Postdoc position in computational neuroscience and vision at Ecole Normale Superieure, Paris and/or at INRIA Sophia-Antipolis (near Nice)

[Romain Brette]

Postdoc position in computational neuroscience and vision at Ecole 
Normale Superieure, Paris and/or at INRIA Sophia-Antipolis (near Nice)

Figure-ground separation with networks of spiking neurons.

Extracting the figure from the ground in a visual scene is one the 
fundamental problems that the visual system faces. As early as in the 
primary visual cortex, the responses of neurons are influenced by 
whether they belong to the figure or to the background, with a 
modulation of their firing rates or correlations in their firing times. 
However, the mechanisms by which the visual system solves this problem 
are mostly unknown. Previous modelling attempts showed that it was 
possible, with lateral interactions, to synchronize neurons responding 
to homogenous regions of the image, but these approaches were limited in 
two ways: 1) they provide a proof of principle but their performance is 
limited compared to modern approaches in computer vision (which are 
themselves limited compared to the actual performance of the human 
visual system); 2) they remain far from the physiological level, because 
they are based on oscillator models, which do not display the irregular 
discharges typical of cortical neurons. This project aims at studying 
how more realistic neural networks (of the integrate-and-fire type) can 
solve the segmentation problem, by relating the dynamics of these 
networks with modern approaches in computer vision.

The most realistic networks for which we have analytical tools today are 
sparsely connected random neural networks. In some regimes, the neuron 
models in these networks display irregular spike trains, and the 
population rate can oscillate at a fast frequency (higher than the 
firing rate of the neurons in the network). Analytical tools have been 
developed to analyze the firing rates of the neurons and the 
characteristics of the global oscillations, but to our knowledge they 
have never been applied to the study of perceptive tasks. In this 
project we propose to apply these techniques to the problem of image 
segmentation, by showing that the network tends to minimize an energy 
that we will relate to energy-based formulations developed in computer 
vision. The latter allow us to express the segmentation problem in a 
Bayesian framework as the one of maximizing the probability of observing 
the image, according to an internal model of statistical homogeneity, 
with the possibility of integrating in a natural way different types of 
information, such as intensity, color, texture and movement.

We are looking for a postdoctoral fellow with expertise in applied maths 
or physics, and experience in programming. Knowledge of 
neurophysiological models is a plus but is not required.

The candidate will work for 1 year in the Odyssee Lab at the Computer 
Science Department of Ecole Normale Supérieure in Paris and/or at INRIA 
Sophia-Antipolis, near Nice on the French Riviera. The position is 
funded by a Marie-Curie Research Training Network (Visiontrain). The 
candidate must not be a French national.

Candidates should send a CV and the address of two referees to Olivier 
Faugeras (Olivier.Faugeras at sophia.inria.fr) and Romain Brette 
(Romain.Brette at ens.fr).