APPLICATION PROCEDURE

The review of applications will begin immediately.

Required application materials:

* complete scientific curriculum vitae
* copy of Masters or Diploma certificate
* copy of PhD certificate, if applicable
* statement of research interests and achievements
* at least two letters of reference
* proof of proficiency in English (e.g., TOEFL or similar)

Please send electronic files and scanned-in versions of documents (all in PDF format if possible).
Files should be compiled into a ZIP archive.

Applicants are asked to apply directly to up to three of the projects
listed below. Write a single application and address it to the
principal investigator(s) named as contact for the corresponding
project(s). As subject line, please use “Bernstein: Application for
PhD Position” or “Bernstein: Application for Post-doc Position”.

DESCRIPTION OF PROJECTS

Development of an Integrated System for Visual Recognition
(Core project)

In the core project we will integrate algorithms for visual perception
and learning. Research will profit from the fact that many existing
systems of perception and learning in vision solve complementary
problems. The combination of different such systems will enable the
development of a vision system with far-reaching capabilities.  The
essential challenge in the core project is to design, implement and
evaluate a common software architecture that allows for the
integration of very diverse visual sub-functions and their
learning-based cooperation in order to solve complex vision tasks.
For further information see: http://fias.uni-frankfurt.de/bernstein

PIs:     Rudolf Mester, Christoph von der Malsburg, Jochen Triesch,
Cornelius Weber, Jorg Lucke

Contact: for this project, contact _all_ three addresses:
mester@vsi.cs.uni-frankfurt.de,
malsburg@fias.uni-frankfurt.de,
triesch@fias.uni-frankfurt.de

Autonomous Learning in an Infant-Like Active Vision System

We will develop an active vision system that autonomously learns
to perceive the world around it. An existing anthropomorphic robot head
capable of fast saccade-like eye movements will be used. In close
collaboration with some of the other projects, the robot will be given
learning capabilities including attentional mechanisms and curiosity
drives. The robot will learn to control its gaze and learn both
low-level (stereo, motion) and high-level (shapes, objects, people)
representations and predictive models in an autonomous fashion.

PIs:     Jochen Triesch, Cornelius Weber, Christoph von der Malsburg
Contact: triesch@fias.uni-frankfurt.de

A Learning Visual Sensor System for a Mobile Platform

This project addresses the demonstration of learnt and continuously
improving visual perception used on a mobile robot which shall safely
navigate in an unknown indoor environment, detect and identify
obstacles and moving objects in the scene. The emphasis is not on the
navigation capabilities but on perception: the project emphasizes the
autonomous learning of motion and near-field environment perception
capabilities under egomotion, considering the conjunction of
perception (vision) and action (motor signals).

PIs:     Rudolf Mester, Hanno Scharr
Contact: mester@vsi.cs.uni-frankfurt.de

Cooperative Neural Learning Approaches in
a Multi-Camera Visual Surveillance Scenario

We plan to develop a system which exhibits autonomous learning of
convergent cooperative processing of visual information in a large
multi-camera setup which is arranged over an extended area. The
demonstrator will show prototypically that a complex network of visual
sensors can learn about the geometric and photometric interrelation
between shared cameras, learn about the appearance and behavior of
people, and ultimately learn about usual and unusual events in an
autonomous fashion. This will be achieved by combining statistical
methods and neural control and communication strategies.

PIs:     Rudolf Mester, Jochen Triesch
Contact: mester@vsi.cs.uni-frankfurt.de

Cue-Integration in Large-Scale Multi-Modal Sensory Systems

Integrating various sensory modalities or submodalities is a
fundamental activity of the brain. Often, the brain seems to integrate
sensory signals in a close-to-optimal fashion. But how does it learn
to do so? In this project we will develop computational models to
explain how the brain forms efficient representations for sensory
signals from different (sub-)modalities and how it learns to integrate
them in an optimal fashion at the same time.

PIs:     Tobias Rodemann, Jochen Triesch
Contact: triesch@fias.uni-frankfurt.de

Hierarchical Memory Models

In this project we will develop and investigate hierarchical memory
structures for visual objects. We aim to develop explicit object
representations that can serve as a basis for visual recognition.  The
studied systems learn from examples with no or little supervision.
Probabilistic generative methods and dynamical systems approaches will
be used. A background in mathematical modelling is desirable.

PIs:     Jorg Lucke, Christoph von der Malsburg
Contact: luecke@fias.uni-frankfurt.de

Analysis of Non-linear Dynamical Systems

In this project we will develop methodologies to analyze large
dynamical systems, to serve as a theoretical foundation of the
dynamical system construction of the core vision system.  Methods in
nonlinear system analysis will be used, and extended with the
assistance of numerical calculation to deal with general systems with
less symmetry.

PIs:     Junmei Zhu, Christoph von der Malsburg
Contact: jzhu@fias.uni-frankfurt.de

Generative Models for Learning and Recognition in Vision

Probabilistic generative models represent a state-of-the-art approach
to component extraction. The vast majority of generative models used
for this task assumes a linear superposition of components.  For
visual data this assumption is often violated. In this project we study
a novel class of generative models which are not limited to linear
component superposition and thus well-suited for applications to
visual data. A background in mathematical modelling as provided by
courses in mathematics, theoretical physics etc. is desirable.
For further reading please see: fias.uni-frankfurt.de/~luecke

PIs:     Jorg Lucke, Julian Eggert
Contact: luecke@fias.uni-frankfurt.de

On-Camera Foveated Vision (FPGA Implementation)

The early human visual system compresses data via high visual
resolution at the fovea (gaze center) but low peripheral
resolution. In this project we will implement such foveated vision,
meeting high standards to foster its wider use, e.g. in other
Bernstein projects. An on-camera implementation is desired, as via a
programmable FPGA or via a “smart camera”. The investigation of
learning algorithms for low-level visual processing makes this project
an ideal entry into computational neuroscience.
Further reading: fias.uni-frankfurt.de/~cweber/08WeberTriesch_fovea.pdf

PIs:     Cornelius Weber, Volker Lindenstruth
Contact: cweber@fias.uni-frankfurt.de

Structural Learning of Motion and Depth Estimation

This project aims at developing new processing structures for visual
motion signals which take into consideration information from numerous
different visual processing ‘modules’, such as stereo, optical flow,
texture flow, higher-order covariance analysis etc. The developed
algorithms will use stereo imagery together with motor signals, and
egomotion data provided by an already available mobile robot platform.
For further reading please visit:
http://www.vsi.cs.uni-frankfurt.de/research/stellenhinweise.html

PIs:     Rudolf Mester, Hanno Scharr
Contact: mester@vsi.cs.uni-frankfurt.de

Neural Models of Development of Visual Processing and Memory in Human
Infants

The human visual system learns to perceive and understand the visual
world in a largely autonomous fashion. But how do we develop an
understanding of fundamental concepts of space, time, objects, or
causality? How do we form memories and models of the physical and
social world around us? In this project, we study selected questions
from this area by a combination of experiments with human infants and
the development of computational models. The models shall be rooted in
biologically plausible learning processes and explain the improvement
of infant competence as studied with methods from developmental
psychology.

PIs:     Thorsten Kolling, Monika Knopf, Jochen Triesch
Contact: triesch@fias.uni-frankfurt.de

Modeling the Role of Feedback Signals in Visual Motion Processing

Massive feedback from higher to lower processing areas are a hallmark
of the cortical architecture. This project combines physiological,
anatomical and computational modeling approaches to investigate the
functional role of these feedback connections for motion
processing. On the experimental side, we will use reversible cooling
of the cortex to selectively deactivate higher processing areas and
observe the impact of the (lack of) their feedback signals on
processing in V1. On the modeling side, we will develop models that
aim to explain the functional role of these connections
(e.g. contributing to Bayesian inference) as well as the learning
mechanisms that shape them.

PIs:        Ralf Galuske, Jochen Triesch
Contact PI: triesch@fias.uni-frankfurt.de

–

For more information about the collaborating institutes please see:

- Neuroscience, FIAS, Goethe-University Frankfurt
http://fias.uni-frankfurt.de/neuro

- VSI, Dept of Computer Science, Goethe-University Frankfurt
http://www.vsi.cs.uni-frankfurt.de/research

- MPI for Brain Research, Frankfurt
http://www.mpih-frankfurt.mpg.de)

- Honda Research Institute Europe
http://www.honda-ri.de