Is anything more magical than sight? The main way we perceive our surroundings, and apparently so natural. In reality, so sophisticated. To the extent that, after some thirty years of research, there is no satisfactory model of visual perception. Not those built by mathematicians and computer scientists, specialists in image processing, computer vision and robotics that only work properly in very controlled environments, nor those built by biologists and neuro-physiologists, although they were the founders, in the 1970s, of neuron behaviour modelling. This observation is a core subject for the ODYSSÉE project, launched by Olivier Faugeras in 2002. The €2 million subsidy granted by the European Research Council for his NERVI project will allow him to spend five years developing a comprehensive formal model of visual perception with the fifteen or so PhD students and post-doctoral researchers he will be recruiting. Although the kernel of the project is essentially theoretical, spin-offs may arise long term. In the medical field, this work might overcome certain visual deficiencies such as deterioration in the perception of contours, textures or colours. Technologically, it could culminate in new image processing algorithms, more robust and better suited to real-life environments. What are the secrets to building this future model? "For it to match our ambitions," replies Olivier Faugeras, "it must be able to simulate how our visual system works on more than one scale in time and space."
The model we are going to try to develop is expected to open new horizons and build bridges between research into biological vision – the neuro-physiologists' research field – and research into artificial vision – the preferred field of mathematicians and computer scientists.
The simulated fields will extend down to the microscopic scale, that of individual neurons, and up to the macroscopic scale, that of the cortical networks, and will include cortical columns, groups of a few dozen to a few billion neurons with similar behaviour. In terms of timescales, it means simulations ranging from the millisecond, for a nerve impulse, to decades. "We will be making use of two mathematical mainstays developed in the last five years in the ODYSSÉE project," adds Olivier Faugeras. Firstly, bifurcation theory, which is still developing and to which we will be contributing to model the considerable variety of nerve impulses from individual neurons. Secondly, mean field theory, based on statistical physics, to model cortical columns and networks."
From design to experimental measurement
The team has the rare ability of proficiency in these delicate mathematical tools that make it possible to "summarise" group neuron activity correctly in conceptual terms. This theoretical approach will be compared with experimental measurements such as cortical column activity in primates being shown sequences of natural images, accessible using optical imaging. Likewise, the electrical activity of cortical networks is accessible through brain imaging techniques such as MRI or magneto-encephalography. We shall see!
"It is an interest group served by great communication intelligence."
Guillaume Masson "Dynamics of visual perception and action" at the Mediterranean Institute for Cognitive Neuroscience (Marseille). He has been working with Olivier Faugeras' team for five years.
Dialogue has naturally built up with this team, and a very intuitive discussion has developed between us over the years. We share the same algorithmic issues with complementary skills sets - in biological visual perception, neuroimaging and neurosciences in my team, and in modelling artificial vision in Olivier Faugeras' team. Consequently, the solutions he developed in image processing match the computational principles we have identified in the cortex. More practically, we jointly supervise doctoral theses and we jointly teach courses. Several of his students have even done their biology theses in my team. We have written articles together and we are involved in national and European projects such as Facets.