Title: Natural scene statistics and visual motion processing in humans and non-human primates
Supervisor: Guillaume MASSON
Laboratory: Institute of Neurosciences, Timone (INT)
State of the art: The mammalian visual system is tuned to process the complex, high-dimension statistics
of natural scenes. For instance, neuronal responses to natural inputs are more finely tuned, more
temporally precise and more sparse. A major challenge is to understand how such complex information is
integrated to measure perceptual quantities such as object motion in a crowded environment. From a
theoretical perspective, this implies to understand how the cascade of cortical processing steps extracts the
relevant dimensions through linear (e.g. filtering) and nonlinear (e.g. gain control, contextual modulation)
mechanisms. From an experimental perspective, the challenge implies to design behavioural tasks that can
dissect out these mechanisms. The classical approach is to compare classic, low-dimension stimuli such as
gratings or dots with natural images. However, it is almost impossible to parametrize natural scenes in a
useful way. We [1], and others [2], have designed and begun to use naturalistic stimuli as static or dynamic
textures of which mean and variance along each parameter space is manipulated to demonstrate how
sensory systems sense and represent information. We have recently shown in Nature Neuroscience3 that
richer textures drive more accurate and reliable ocular tracking responses. This is the first demonstration
that naturalistic inputs processing can be probed using sensorimotor tasks.
Objectives: Our objective is to demonstrate that in human and non-human primates, estimating direction
and speed of a moving object requires to integrate information from the multiple spatial and temporal
scales existing in natural scenes. We will investigate how different spatiotemporal frequency channels are
nonlinearly combined to drive faster and more precise tracking eye movements. In particular, we will
extend our previous study [3] by studying how richness of visual motion information reduces the variability
of motor responses, within and across trials. We will also study how the geometric organization of the
texture input impacts performance, opening the door to understanding the link between low-level motion
processing and the perceptual organization of complex scenes.
Methods: The approach is based on probing low-level visual processing through the dynamics of
reflexive eye movements. Ocular following responses exist in both humans and non-human primates and
over the last 15 years we, and others have shown how they can probe low-level motion mechanisms. The
PhD will work in non-human primates to conduct behavioural experiments using random phase textures
and recording eye movements with the scleral search coil technique. Behavioural performance will be
assessed through kinematics (acceleration, latency…) as well as through detection and discrimination
tasks (i.e. comparison between responses for two stimuli).
Expected results: A first step will be to reproduce the main observation made in humans: richness of the
inputs improves precision and decreases variability. Then we will characterize the nonlinear interactions
between spatiotemporal channels underlying speed estimation through motion energy and their temporal
dynamics.
Feasibility: This project is supported by an ANR project (SPEED) that started in October 2013. The PhD
student will work in close collaboration with a PD fellow working in humans as we expect the first series
of experiments to be identical in both species. A non-human primate set-up is available and a research
assistant (F Barthelemy) will daily supervise the lab work. The InViBe team has strong expertises in
modelling and electrophysiology so that the project will open the door for collaborative work in texture
synthesis and speed processing modelling with Laurent Perrinet and recording of population activity
evoked with dense, naturalistic textures with Fred Chavane.
2 Freeman et al. (2013) Nature Neuroscience