Connectionists: postdoctoral position, call for applications

[Bernard Girau]

A postdoctoral position may be available at INRIA/Lorraine, Nancy, France.
The deadline for submitting applications is very close: 14 of July 2006.


------------------------------------------------------------------------ 
-------


               INRIA-LORRAINE

        INRIA 2006 POSTDOCTORAL FELLOWSHIP PROPOSAL



------------------------------------------------------------------------ 
-------


City :
Nancy, France

Title of Research Project:
Feedback interactions for the visual perception of motion by neural networks

INRIA team in which the research would be effected:
Cortex

Fellowship supervisor:
Bernard Girau, Assistant professor at University Nancy 2, Cortex project


Description of Research:

Our connectionist research group carries out numerous projects about
the use of neural networks for autonomous systems (autonomous
robotics, embedded processing of physiological signals, ... etc).
These works take advantage of the elementary and massively distributed
computations of connectionist models. They include for example
perceptive pre-processing modules (vision, olfaction) or procedural
control, that aim at interacting autonomously in a real-time and
embedded way.

The connectionist paradigm has to be the foundation of the autonomy of
the processings we develop. Therefore their specific very fine-grain
massive parallelism has to be fully exploited, where computation units
take place in a very dense information stream. The organisation and
the interpretation of such a stream is not a simple problem. Therefore
our research group takes advantage of neuro-physiological knowledge to
define connectionist architectures that may perform particularly
complex cognitive tasks.

A PhD student of our team has developped a modular, highly local and
distributed connectionist model for the visual perception of motion
[1]. The structure of this model derives directly from the course of
the optical flow in the human brain. It is mainly using a dense set of
excitatory and inhibitory local lateral interactions.  A first module
performs a low-level spatio temporal filtering to detect speeds.  A
second module handles this information through an inhibition mechanism
inspired by the organization of cortical area V1 as a set of
orientation columns, so that more coherent motion areas are extracted.
Finally, a last module focuses on a single moving object and it
performs object tracking in natural image sequences [2].

This model offers promising results in real
environments. Nevertheless, numerous improvements are required. The
intermediate module (coherent motion areas extraction) currently
stands as the most difficult aspect to handle, and its performance is
still beyond our expectations.  Moreover, our model is not able to
extract egomotion from image sequences, nor to make object tracking
specialize for predefined kinds of motion.  Finally, numerous
capacities of the focus attention module are not exploited, such as
its ability to track several targets simultaneously or to change the
tracked target on demand.

These limits of our connectionist model for motion perception are
currently mainly caused by a lack of modeling of an essential aspect
in the neurophysiological pathway of visual signals: it should take
into account the dense feedback interactions that exist between the
involved cortical areas.

The goal of this subject is to improve our connectionist model for
motion perception thanks to the introduction of feedback interactions
between the different modules.  This work will take advantage of
existing studies about such interactions in the human brain.  The main
part is the insertion of the information extracted by the focus
attention module inside the excitatory/inhibitory mechanism that
extracts coherent motion areas. Then this work will deal with the
backpropagation of coherent area informations to the detectors of
local motion, so as to improve the robustness of our model.

[1] C. Castellanos-Sanchez and B. Girau. Dynamic pursuit with a 
bio-inspired neural model.
In ACIVS 2005, volume 3708 of LNCS, pages 284--291, 2005.

[2] J. Vitay, N. Rougier and F. Alexandre. A distributed model of visual 
spatial attention.
Biomimetic Neural Learning for Intelligent Robotics, 2005.


Desired profile of candidat :

We aim at recruiting a post-doctoral fellow having a
strong background in connectionism, preferably applied to
visual perception. This work would also take advantage of
skills in software engineering and in computational modeling.

Expected duration of fellowship:
12 months

Contact:
Bernard Girau
E-mail : girau at loria.fr
Tel : +33 3 83 59 20 58