PD Dr. Rolf WÜrtz, Theory of Neuronal Systems, Institute for Neural Computation

Rolf Würtz

Research Programme:

The driving force behind my research is the intellectual challenge to arrive at a better understanding of human perception. This goal can be convincingly reached by presenting machines that can simulate perception-like performance under realistic circumstances.

The logical constraints encountered during development also apply to living brains. Such machines will obviously be of outstanding commercial interest, and their development should be accompanied by solving practical problems wherever possible. They will also bridge a serious gap in current artificially intelligent systems by providing them with a means of assessing results of reasoning by direct inspection of the environment. Therefore, the active acquisition and interpretation of information is a central goal. More concretely, my current research interests are as follows:

  • Further research on the visual correspondence problem.
  • Extension of my methods for object recognition to further invariances and object classes.
  • Steps from object recognition to scene analysis.
  • Integration of flexible matching techniques and fast associative memories.
  • Analysis of the four dimensional Gabor phase space of natural images.
  • Integration of different feature detectors to a robust object description, which is in turn suited for matching into new images or image sequences.
  • Vision-based robot calibration.
  • Autonomous learning of object representations for recognition and grasping.
  • Integration of visual and tactile information for object handling.
  • Automatic interpretation of human emotions and gestures.
  • Steps towards a solution of the symbol grounding problem.

Practical/method courses

2 week practical training "Visual object recognition" held annually in September. Enscription is required in June.

In this lab exercise image processing will be studied with practical examples from visual object recognition. Each day starts with an introduction, which presents the theoretical foundations and the problems to be solved. Each of the subsequent practical parts uses a ready made computer program, which can be used to solve the problems applying minor modifications. Concrete topics are:

  • Introduction to Linux
  • Introduction to C++ programming
  • Image acquisition and coding, simple pixel operations
  • Filtering, template matching
  • Frequency space, Fast Fourier Transform
  • Wavelet coding of images, reconstruction
  • Object representation, labeled graphs
  • Recognition and classification

Recommended prior experience: It is recommended to have attended one of the lectures "Vision in Man and Machine" or "Perception in self-organizing Neural Networks". Basic knowledge of signal theory and programming is helpful.

Literature: B. Jähne: Digital Image Processing, Springer. Handout "Visual object recognition"

Please note that currently this course including course material is given in German. Therefore, for students not speaking German enscription no later than June 30 is mandatory.

Lecture and exercises:
Network Self-Organization in the Ontogenesis of the Visual System The course gives an introduction to neurobiological phenomena, the concepts and the methodology of self-organizing systems and will be based largely on systems of non-linear differential and integro-differential equations. Applications include the emergence of ordered connections structures in the visual system (retinotopy, orientation columns, ocular dominance strips). All relevant terms and methods are introduced, but elementary mathematical knowledge is a prerequisite for participation. In the exercises, models of self-organization are simulated on workstations.

Lecture and exercises:
Perception in Self-Organizing Neural Networks Visual and auditory perception pose a couple of fundamental problems like scene segmentation, invariant object recognition, scene representation and analysis. This course gives a brief introduction to these problems discusses some classical approaches on the basis of neural networks. After a critique of these approaches the Dynamic Link architecture will be introduced as an extension of Neural Networks. It rests on processes of rapid network self-organization. Much space will be given to the solution of fundamental problems of perception on this basis. In the exercises, the relevant models are simulated on workstations, among them a face recognition system.

Lecture: Vision in Man and Machine
This lecture treats aspects of vision from the computer science, psychophysics, and neurobiological point of view. Starting from a phenomenology of human vision, biological and psychological basics, the foundations of image processing are developed, finally leading to more advanced concepts like multi-scale and wavelet analysis, and algorithms for face and object recognition. The lecture is suited for students of computer science, biology, physics, engineering, and psychology.

Lecture and exercises: Artificial Neural Networks
This lecture presents standard algorithms and new developments of Artificial Neural Networks, their functioning, application domains, and connections to more conventional mathematical methods. Examples show the potential and limitations of the methods. Supervised as well as unsupervised learning methods are introduced.

Seminar: Organic Computing
Living organisms are capable of producing goal-directed processes by self-organization. It can be assumed that general principles underlie this sort of organization, from the amoeba up to vertebrate brains. What do we know about these principles? How can the organization of the genome be uncovered? What is the contribution of neuroscience? Is it a good idea to organize a computer organically rather than by algorithmic control? In the seminar the algorithmic and the organic approach to problem solving will be discussed and an organic view on software technology will be motivated. Organizational principles of natural and artificial information processing systems will be explored, and common features (like modularity and robustness) will be uncovered. Finally, some artificial information processing systems will be presented, that are built on these findings. As an interdisciplinary course this seminar invites students of Biology, Electrical Engineering, Computer Science, Mathematics, Medicine, Philosphy, Physics, and Psychology. Successful participation will be documented with a certificate that can be honored by the respective study commissions.

Methods Course: Visual Object Recognition
In this lab exercise image processing will be studied with practical examples from visual object recognition. It is structured into successive units of 2-3 days. Each unit starts with an introduction, which presents the theoretical foundations and the problems to be solved. Each of the subsequent practical parts uses a ready made computer program, which can be used to solve the problems applying minor modifications.
Concrete topics are:

  • Introduction to Unix
  • Introduction to C++ programming
  • Image acquisition and coding, simple pixel operations
  • Filtering, template matching
  • Frequency space, Fast Fourier Transform
  • Wavelet coding of images, reconstruction
  • Object representation, labeled graphs
  • Recognition and classification