In January 1990, led by the Simulation Program Informatics in the Netherlands (SPIN) of the Ministry of Economy, a national study program on neural networks started. This study is carried out by SNN. The objectives of SNN are to promote the basic and application studies on neural networks and neural information processing in the biology and artificial system areas, to identify prospective areas for applications and to transfer expertise on neural networks to industry applications.
SNN coordinates most of the studies on the neural networks conducted in the Netherlands. In these programs they study robotics, vision and pattern recognition and perception systems as well as neural information processing methods for various applications. SNN constantly monitors the latest researches and developments in this active field and leads new projects in areas industries have great intereste in. SNN is working jointly with industries on resolving problems of specific industrial applications using neural networks. At present, the departments of neural network study in University of Nijmegen, Utrecht University, University of Amsterdam, Delft University of Technology (TUD) and University of Groningen are participating in SNN. Further details may be obtained through http:/www.mbfys.kun.nl/snn.
RWC studies are undertaken by the study group of Amsterdam University, Utrecht University and Nijmegen University.
The University of Amsterdam group is engaged mainly in sensing function of target tracking and robot navigation for a robot to be implemented in the real world. The group in Utrecht University traditionally has forte for developing multiresolution models for robust vision. The University of Nijmegen group has conducted many theoretical studies on neural networks. They also boast the tradition of experimental studies on the brain.
These groups have experiences in conducting studies at SNN on a robotic systems comprising multiresolution vision systems and path planning. The neural networks learn control and path planning tasks. The videos are available which explain this system.
SNN is participating in active cognition and perception project of RWCP. The target of this project is to develop new theories, methods and implementations for perception, inference, and action plan under the complex and dynamic multi-sensor environment. The results of the study will be applied to robotics and multimedia aplication. Methods for the Active cognition and perception should be applicable to other areas such as industrial processes, medical image processing, decision-making support, financial analysis, language and script recognition.
Presently SNN is engaged in important basic problems of the real world computing as follows:
Probabilistic Knowledge Representation
Conventional rule-based systems that rely on pure logic can not handle uncertain (incorrect, incomplete or inconsistent) data. Besides trivial cases, this especially matters in the application to the real world computing that will not be able to get complete knowledge. The key issue is to design systems that are semantically correct and computationally efficient. Also, these system must have an ability to learn from data and to incorporate structural knowledge on the domain. We propose to develop methods that link domain information with the adaptive method (e.g. neural network) in the framework of the probability theory. These methods will be demonstrated in the application to medical diagnosis.
Learning in Changing Environment
The stability of representing learning is an extremely important problem for learning in a changing environment. What is the criteria for a system to judge whether or not a certain representation is still appropriate in the current environment? To what degree of adaptability is required for the system? We succeeded in developing a general theory on the actions of a learning system under a changing environment. We are now developing a robust learning algorithm based on this theory.
Integration of Reactive Actions and Planned Actions
There is inconsistency between command level and execution level in many robot and non-robot systems. While command level is tend to be formulated in high-level symbolic languages, execution level is tend to be formulated in low-level device control commands. For the present, the transition between these levels is done manually for each application and not understood in general terms. It is critical to design a universal method that will integrate the command level (planned actions) and the execution level (reactive actions). We plan to develop appropriate intermediate representations using an office mobile robot as an example.
Multi-resolution Knowledge Representation
Pre-processing such as the extraction of image structure using a low-level vision processing operators is essential for high-order and high-level robot vision such as perception and learning. The scale space theory has been developed recently from computer vision studies on pre-attention vsions or unconcerned visions. The scale space theory explicitly takes into consideration multi-scale hierarchical structure of images and is extremely robust against images that contain a lot of noise. We suggest that this mathematical framework to be applied to extract image structures required for input to the proposed neural networks and robots. We plan to demonstrate our method through the design of a 3-dimensional, visual mulit-resolution image editor and a feature detection system for object recognition in the context of robotics. The results may be applicable equally to earthquake prediction, remote sensing, video image processing (for amusement industry), monitor system, and visual inspection.
 a)A Road with a Cypress under the Starlight by Vincent van Goah (Rijksmuseum Kroeller-Mueller) |  b) A filtered image closes interrupted lines. This is useful for processing finger print and other striped patterns. |
*J.A. Weickert,B.M. ter Haar Romeny, M.A. Viergever: Conservative image transformation with restoration and scalespace properties. Proc. 1996 IEEE Int. Conf. on Image Processing(ICIP-96, sept. 16-19, 1996, Lasusanne)