Wolfgang Eppler

New control system aspects for physical experiments

New control system aspects are introduced for the design of slow control systems for physical experiments. Mainly, they are based on a comprehensive usage of XML technologies. A second paradigm for future control systems is the consequent usage of the model-view-controller (MVC) pattern. Main aspect of all (hardware and software) system components is the use throughout of standards for interfaces, protocols and architecture. In particular, the software is based on the unifying XML specifications XSchema, XPath, XQuery, XLink, and OPC XML. A first application of these technologies is the KArlsruhe TRItium Neutrino (KATRIN) Slow Control System (KSC) of a neutrino experiment at the Forschungszentrum Karlsruhe. Main characteristic of KSC is its homogeneous structure with its code being spread over several subsystems. Implications of this distributed method are system stability, independence of subsystems and fast and comfortable maintenance.

Neural chip SAND/1 for real time pattern recognition

The neural chip SAND/1 (Simple Applicable Neural Device) was designed to accelerate computations of neural networks at a very low cost basis. Low cost means that only few peripheral chips are necessary to use the neural network chip in applications. The design criterions of the chip are described. A PCI-board was developed and a VME-board is under construction to facilitate applications. At present, time series prognosis, multidimensional curve fitting, pattern recognition in gamma-shower detection and mammography are under evaluation. The last application is described in more detail.

The SAND Neurochip and Its Embedding in the MiND System

The system MiND (Multipurpose integrated Neural Device) is a tool for the development of artificial neural network applications which integrates hardware and software components. It includes a PCI neuro-board with up to four SAND (Simple Applicable Neural Device) neuro-chips. The neuro-board accelerates feedforward networks, Radial-Basis-Function networks, and Kohonen feature maps. There are several simple to use software layers for exploiting the neuro-board. At the bottom, there is the drivers C interface. Secondly, a number of C++ network classes are built on the C-drivers. Thirdly, comfortable simulators with graphical interfaces base on the C++ classes. These stein from a pool of “predefined” simulators provided by the MiND system. Each simulator is constituted by a network definition written in the neural network description language CONNECT, and by an interface definition script. The interface definition is based on a C++ network class generated from the CONNECT definition, and on abstract graphical user interface classes. A user can develop own network or interface definitions. Also, the C++ network classes can be exported for integration into custom applications.

Ultrasound Computer Tomography, Distributed Volume Reconstruction and GRID Computing

At Forschungszentrum Karlsruhe the demonstrator of a new medical imaging system based on 3D ultrasound computer tomography will be available at the beginning of 2005. This system requires not only the recording but also the processing of very large datasets of about 3.5 GBytes. On a single PC the reconstruction of a high resolution volume will take up to several weeks. However the reconstruction process can be parallelized. Our long term goal is to accelerate the reconstruction process significantly by the use of GRID technologies. In this paper we present our approach to parallelize the reconstruction by A-scan-decomposition using Java RMI and Matlab.

High speed neural network chip for trigger purposes in high energy physics

A novel neural chip SAND (Simple Applicable Neural Device) is described. It is highly usable for hardware triggers in particle physics. The chip is optimized for a high input data rate (50 MHz, 16 bit data) at a very low cost basis. The performance of a single SAND chip is 200 MOPS due to four parallel 16 bit multipliers and 40 bit adders working in one clock cycle. The chip is able to implement feedforward neural networks with a maximum of 512 input neurons and three hidden layers. Kohonen feature maps and radial basis function networks may be also calculated. Four chips will be implemented on a PCI-board for simulation and on a VWE board for trigger and on- and off-line analysis.

Volume Reconstruction of Clustered Micro-Calcifications in Mammograms

Breast cancer is one of the leading causes for death of women in the western world. Detection of anomalies early pathological stages and immidiate treatment are essential for successful cure. Early stages of breast cancer are indicated by the occurance of microcalcifications [2]. About 30% of all tissue changes show clustered microcalcifications. Their shape and spatial arrangement are of high diagnostic value [3]. In many cases it is possible to distinguish benign and malignant tissue changes on the base of the shape and spatial arrangement of its contained microcalcifications [4]. To recognize the positions of microcalcifications in the female breast, radiologists have to detect them in mammograms and have to analyse their spatial relationship just on the base of one or two views/mammograms. The process of recognition and reconstruction of clustered microcalcifications requires a good expert knowledge and a high abstract imagination capability. Therefore it is very useful to recogize microcalcifications automatically in mammograms and to present their spatial relationship in an animated 3D modell.

A Low Cost Computer Assisted Mammography Workstation

Conspicuous features in X-ray mammograms may be hard to detect in early cancer stages and may be missed even by experienced radiologists. In principle, mammograms should be regarded independently of the first evaluation by a second expert. Computer assisted diagnosis can play the role of the second expert in regarding X-ray mammograms. Automatic detection of microcalcifications and masses will support the radiologist to reduce the probability of overlooking subtle tissue features, visualization of the three-dimensional arrangement within clustered microcalcifications will help to discriminate benign and malign cases.

Symbolic Learning in Connectionist Production Systems

The paper discusses an approach to combine classical artificial intelligence with connectionism. The crucial point with this intention is the implementation of symbols in neural networks. Several proposals are mentioned in [2]. The next important thing to be examined is the increase or modification of knowledge in a connectionist system. In this paper two types of learning are introduced: subsymbolic learning by experience and symbolic learning from linguistic rules. Symbolic learning requires two regions of processing, the language region and the signal region, both of them being coupled with associative links. Symbolic rules are processed sequentially in several cycles, and affect mainly the language region by also influencing the signal region, whereas subsymbolic rules are local to each of the regions and their consequences can be concluded in parallel. In a combined system with both kinds of rules subsymbolic learning and subsymbolic rules are the basis for symbolic learning and the application of linguistic rules. First results with a prototype system are introduced.

A DIGITAL SIGNAL PROCESSOR FOR SIMULATING BACK-PROPAGATION NETWORKS

A digital signal processor for simulating multi-layer perceptrons with back-progagation as learning algorithm (BP-DSP) is introduced. The objective of this approach is rather to show that huge networks can be simulated under real time conditions than to develop a simulation machine for a special hardware. Some experiments are described to determine the data width and the interpolation of the sigmoid output function. Examples show the interdependence of the algorithms implemented. The paper concludes with a comparison of speed for some simulation machines.

Entwurf einer integrierten Schaltung zur Beschleunigung von Koordinatentransformationen mit einem Silicon Compiler

Fur die Losung vieler Probleme im CAD/CAM Bereich — insbesondere in der Robotik — ist der Entwurf besonders schneller Schaltungen notwendig, um die erforderlichen kurzen Reaktionszeiten zu erreichen. Wichtige Schwerpunkte unserer Arbeit sind in diesem Zusammenhang die Entwicklung einer geeigneten Architektur fur die Realisierung der Algorithmen und die Integration der entwickelten Schaltungen als Coprozessor in ein bestehendes Rechnersystem. In diesem Beitrag wird der Entwurf einer anwendungsspezifischen integrierten Schaltung fur die Berechnung trigonometrischer Funktionen, sowie die Belegung von Denavit-Hartenberg Matrizen und deren Multiplikation vorgestellt. Benotigt werden diese Funktionen bei der Koordinatentransformation, die bei der Steuerung eines Roboterarms eine wichtige Rolle spielt. Fur die Umsetzung der Algorithmen in eine hochintegrierte Schaltung wurde ein Silicon Compuer verwendet.