Inference in Visual Behaviours

Movements of our eyes as well as our perception are primarily based on the ability of the visual system to estimate accurately the location and movement of objects in a cluttered environment. The INVIBE team aims at better understanding how such information is extracted (visual motion processing), how they are filtered depending upon the behavioral constraints (attention, decision-making) and how they are transformed into motor commands controlling saccades and pursuit and coordinating them (oculomotor control). The team combines behavioral (psychophysics, eye movement behaviors) and physiological (single & multiple units recordings electrophysiology, pharmacological perturbations, MRI/TMS) approaches in both humans and monkeys. Neuronal and behavioral data are also simulated using mathematical models based on Bayesian inference theory and implemented in artificial neural networks.

More specifically, the main research topics in the team are:

  • Visual motion processing for perception and action: We investigate how the visual motion system extracts and selectively integrates local information about object motion about how these representations are used to control reflexive or voluntary eye movements (pursuit) and perception. We combine psychophysical and behavioral methods targeting low-level visual motion processing (local speed and direction processing) and integration/segmentation mechanisms to optimally reconstruct the velocity of, and only of, the target of interest. Comparing perceptual and motor performances allow us to have access to different, context-dependent computational rules and different time scales (e.g. fast detection/discrimination, slow multi-stability). These rules are simulated using dynamical models (Bayesian models, dynamical ring attractor networks). One original approach is to use naturalistic stimuli mimicking the statistical properties of natural images while being still fully parametrically-controlled.

Principal investigators: Guillaume Masson, Anna Montagnini, Laurent Perrinet

JPEG - 66.8 kb

Figure: Main cortical processing steps involved in extracting, integrating and segmenting local information about an object moving in a cluttered environment. These computational steps are dynamical since they are embedded in recurrent networks implementing context dependent, adaptive rules based on the statistical properties of the images (from Medathati et al., International Journal of Computer Vision 2016)

  • Neural bases of oculomotor control: How the brain controls movements of our eyes directed towards a moving goal ? We study on both human and monkeys the nature of visual signals about target location/motion are transformed into motor signals used to coordinate pursuit and saccadic eye movements when tracking simple, continues target motion trajectories as well as complex trajectories with occluded portions or change in direction, speed… In monkeys, we investigate the respective roles of the multiple neural signals flowing between ponto-cerebellar, cerebello-reticular pathways as well as cortical and cerebellar inputs to the superior colliculus in the control of orienting gaze. To do so, we use causal approaches based on local, reversible perturbations (local drug injections, micro-stimulation) applied to different nodes (superior colliculus, brainsteam nuclei) of the complex but versatile oculomotor system.

Principal investigators: Laurent Goffart, Anna Montagnini, Guillaume Masson

  • Acting and perceiving also use prediction: Eye movements, as well as perception, are not only based on current information. Our brain has a striking ability to estimate statistical regularities from target motion trajectories and to use these to optimize motor control by reducing processing delays or by anticipating target motion onset and velocity. We study how probabilistic neural representations of target motions are learned, on which time scales and with which context dependent computational rules. At theoretical level, such probabilities learning is modeled as a Bayesian network where the decision is continuously updated by integrating both new sensory information and accumulated history forming a dynamical Prior distribution. We investigate how the dynamics of these optimal probabilistic representations shape eye movements and perception. We start also investigating their neural basis in humans using fMRI and trans-cranial magnetic stimulation (TMS).

Principal investigators: Anna Montagnini, Laurent Perrinet, Laurent Goffart, Guillaume Masson

  • Decision-making and attention:  Sensory systems receive a continuous flows of dense and complex information from the peripheral sensors. To optimally adapt our behaviors to the environment and the behavioral contexts ones need to process and represent only relevant features (a friend in a crowd, a predator hidden in an high grass field), such information must be compared to internal representations of what ones are looking for. Attention filters information in an adaptive way, taking into account the behavioral and environmental context. Therefore, attention eases perception of expected features, shaping decision-making mechanisms based on the features. We study how prefrontal (PFC/FEF) and posterior parietal (PPC/LIP) cortex interact with extra-striate visual areas (V4, MT) in order to control and orient visual attention, to integrate and bind relevant visual information (spatial and feature information) and take perceptual decisions. We combine psychophysical methods in humans and monkeys with multiple- and single-unit recordings in macaque PFC/PPC and V4/MT areas. To demonstrate their causal role we selectively perturb these neuronal populations using electrical micro-stimulation or pharmacological inactivation.

Principal investigators : Guilhem Ibos, Anna Montagnini

  • Dynamics of inference: computational neurosciences and artificial intelligence. Investigating sensory processing mechanisms allows to define computational rules that can then be implemented in artificial models. More specifically, we investigate how the brain builds a probabilistic estimate of specific target motion properties (e.g. its trajectory) from uncertain and ambiguous local motion information. Such integration takes advantage of a priori knowledge about the statistical properties of our visual world and the displacement of its objects. The probabilistic formalism used in our studies allow to elucidate the relationship between the dynamics of neural computation and the precision of both the sensory measurements and the Prior distributions. These probabilistic models can then be implemented in artificial neural networks and therefore be used when building the next generation of artificial systems as for instance embarked sensory or cognitive systems (robots, autonomous vehicles).

Principal investigators : Laurent Perrinet, Anna Montagnini

JPEG - 82.8 kb

Figure : A dynamical estimate of the trajectory (i.e. its position and velocity) of a moving target (in red). Such estimate is plotted as a function of time (from left to right) and is based upon a diffusion mechanism of measurements precision. Such diffusion allows cancelling processing delays and therefore to represent the target position in real time. By comparison, a second, flashed target (green) will not lead to diffusion mechanism, and therefore no prediction mechanism such that its perceived location will be delayed. Despite the fact that red and green targets are located at the same position at the time of the flash, the moving targe twill be perceived as being ahead of the static, flashed green target, corresponding to the flash-lag effect. From Khoei et al. PLOS Computational Biology (2016)

The INVIBE team is composed of 4 permanent CNRS researchers

  • Guillaume Masson (DR1 CNRS) has studied for 20 years the mechanisms of visual motion processing for both perception and eye movement control. He has published more than 70 research articles, including 8 in top-ranking journals such as Nature, Nature Neuroscience, PNAS et Current Biology and several review articles on biological and artificial vision. He is the founder of the Institut de Neurosciences de la Timone (2012) and its current Director.
  • Anna Montagnini (CRCN, CNRS) is interested in human visual and oculomotor systems as models of how human being makes decision and how such decision-making mechanisms are strongly influenced by the uncertainty of visual and non-visual information. Her work uses both empirical studies in human participants and modeling behavioral performances. She has published more that 20 articles in international journals (J Vision, Nature Neurosci...).
  • Laurent Goffart (CRCN, CNRS) has studied for 20 years the neural mechanisms underlying gaze orientation in non-human primates.. In particular, he is interested in the role of brainstem nuclei and the cerebellum. He dissects their role using reversible inactivation and microstimulation in behaving monkeys. He has published more than 30 articles in high-standard journals (Science, J Neurosci..)
  • Guilhem Ibos (CRCN, CNRS) has been working since 10 years on cortical mechanisms of attention and decision-making. He records neuronal activities simultaneously in 3 cortical areas, while the monkeys perform complex tasks. He has published more than 10 papers, in particular in Neuron and ELife.

Research training:

InViBe hosts Master and PhD students, as well as post-doctoral fellows. PhD students are enrolled in the Integrative and Clinical Neurosciences PhD Program (Aix-Marseille Université, ). PhD students and post-docs are funded by Higher Education and Research fellowships and national and international grants. InVIBE coordinates the PACA Marie-Curie ITN network. The team welcomes foreign students (Master/Erasmus, PhD, post-doc) and currently 2/3 of INVIBE fellows are foreigners. We recruit students with a background in neurosciences, experimental psychology, computer sciences or applied mathematics.

A list of proposed training projects are listed here. Students are encouraged to seek information about open job positions as well as training opportunities : training.

CNRS logo université Aix Marseille logo | Sitemap | Terms & Conditions | contact | admin | intranet | intcloud |