Udo Seiffert

Multi-Dimensional Self-Organizing Maps on Massively Parallel Hardware

Although available (sequential) computer hardware is very powerful nowadays, the implementation of artificial neural networks on massively parallel hardware is still undoubtedly of high interest, not only under an academic point of view. This paper presents an implementation of multi-dimensional Self-Organizing Maps on a scalable SIMD structure of a CNAPS computer with up to 512 parallel processors.

Directed random search for multiple layer perceptron training

Although backpropagation (BP) is commonly used to train multiple layer perceptron (MLP) neural networks and its original algorithm has been significantly improved several times, it still suffers from some drawbacks like being slow, getting stuck in local minima or being bound to constraints regarding the activation (transfer) function of the neurons. The paper presents the substitution of backpropagation with a random search technique which has been enhanced by a directed component. By means of some benchmark problems, a case study shows general potential application fields as well as advantages and disadvantages of both the backpropagation and the directed random search (DRS).

Estimating motion parameters with three-dimensional self-organizing maps

Abstract The importance of analyzing moving scenes within the wide area of digital image processing is increasingly high. Although a simple detection of object velocity by neural networks has been considered in previously published papers, an implementation of artificial neural networks using a priori information for motion analysis is still quite rare. This paper shows the benefits from artificial neural networks, and from using a priori information about the contents of the history in the image sequence to improve the accuracy and speed of estimating motion parameters in the cases of distorted or overlapped objects. In the first place, it introduces three-dimensional Self-Organizing Maps (SOM) with two-dimensional input layers.

Three-dimensional self-organizing maps for classification of image properties

The importance of analysing moving scenes within the wide area of digital image processing is increasingly high. Although a simple detection of object velocity by biological models has been considered in previously published papers (A. Tsukamoto et al., 1993; S. Wimbauer et al., 1994; J. Hogden et al., 1993), an implementation of artificial neural networks using a priori information for motion analysis is still quite rare. The paper shows the benefits from artificial neural networks and from using a priori information about the contents of the history in the image sequence to improve accuracy and speed of estimating motion parameters in the cases of distorted or overlapped objects. Firstly, it introduces 3 dimensional self organizing maps (SOM) with 2 dimensional input layers.

Improving Image Processing Systems by Artificial Neural Networks

Document analysis systems require documents in electronic format. An image acquisition and display system for scanning and saving documents is presented, whereby the recognition capability (for example, series-connected OCR systems) is improved by correction components. Components for improving image acquisition, archiving documents and for reducing compression errors during archiving are integrated in the overall solution. The deployed components are suitably trained artificial neural networks. The projected improvements are assessed.

Adaptive three-dimensional self-organizing map with a two-dimensional input layer

This paper introduces an adaptive growing three-dimensional self-organizing map with the special feature of a matrix input layer instead of a one-dimensional vector. First a short description of growing SOMs is given and the fundamental advantages are mentioned. Then an extension of the original SOM from two to three dimensions with a growing feature is presented. By means of a selected example the general behaviour of the algorithm is illustrated.

Classification of 3-D Dendritic Spines using Self-Organizing Maps

This work in progress shows a method for classifying dendritic spines by their shape. Focal points are the extraction of features from three-dimensional spine data and the following classification of the spines. Hence there will be only little reflection of biological aspects of this problem. Feature extraction based on moments and spherical coordinates will be discussed. Furthermore, this paper shows and describes a modified kind of self organizing maps (SOM)), which is used for the classification of the dendritic spines.