Home Home
Christian Goerick, Ruhr-Univ. Bochum (Germany)
Bernhard Sendhoff, Ruhr-Univ. Bochum (Germany)
Werner von Seelen, Ruhr-Univ. Bochum (Germany)
Artificial neural network have evolved from their biologically inspired roots to a well established means to solve a broad spectrum of engineering problems. The embedding into modern statistics has provided the necessary theoretical foundation for challenging engineering tasks, such as advanced real-time image and signal processing. These are exemplary demonstrations for the applicability of this approach to complex information processing. However, the large number of applications must not obscure the fact that there are some major unsolved problems concerning neural networks. There are still no satisfactorily constructive ways to determine the optimal structure (elements as well as organization) or the learning and evaluation dynamics. The ongoing research addresses these problems. In addition to pursuing this direction, one can ask, what other lessons we can learn from biology concerning complex information processing. Our goal in this paper is to sketch a possible way from neural networks to more comprehensive neural strategies.
ic970119.pdf ic970119.pdf
TOP
Rosaria Silipo, DSI (Italy)
Giovanni Bortolan, DSI (Italy)
Neural and traditional techniques have been compared for the particular task of automatic ECG analysis. A large validated ECG database has been used. Statistical methods, neural architectures with supervised and unsupervised learning, and a neuro-fuzzy architecture have been considered. The results from the connectionist approach are always at least comparable with those coming from more traditional classification methods. But the best performances have been obtained by the combination of the connectionist with the fuzzy approach.
ic970123.pdf ic970123.pdf TOP
Dragan Obradovic, Siemens, München (Germany)
Gustavo Deco, Siemens, München (Germany)
This paper provides a detailed and rigorous analysis of the two commonly used methods for redundancy reduction: Linear Independent Component Analysis (ICA) and Information Maximization (InfoMax). The paper shows analytically that ICA based on the Kullback-Leibler information as a mutual information measure and InfoMax lead to the same solution if the parameterization of the output nonlinear functions in the latter method is sufficiently rich. Furthermore, this work briefly discusses the alternative redundancy measures not based on the Kullback-Leibler information distance and Nonlinear ICA. The practical issues of applying ICA and InfoMax are also discussed.
ic970127.pdf ic970127.pdf TOP
Juha Karhunen, Helsinki University of Technology (Finland)
Aapo Hyvärinen, Helsinki University of Technology (Finland)
Ricardo Vigário, Helsinki University of Technology (Finland)
Jarmo Hurri, Helsinki University of Technology (Finland)
Erkki Oja, Helsinki University of Technology (Finland)
In blind source separation one tries to separate statistically independent unknown source signals from their linear mixtures without knowing the mixing coefficients. Such techniques are currently studied actively both in statistical signal processing and unsupervised neural learning. In this paper, we apply neural blind separation techniques developed in our laboratory to extraction of features from natural images and to separation of medical EEG signals. The new analysis method yields features that describe the underlying data better than for example classical principal component analysis. We briefly discuss difficulties related with real-world applications of blind signal processing, too.
ic970131.pdf ic970131.pdf TOPMark D. Plumbley, KCL (U.K.)
In this paper we shall see that neural networks and communications are interlinked in a number of ways, towards the goal of efficient communication of information. One concrete example of this is the use of neural networks to ensure efficient use of communication channels, through connection admission control in ATM networks. In addition, however, efficient communication is also important within a decision making system such as a neural network. Finally we examine what type of neural network solutions are suggested by this approach.
ic970135.pdf ic970135.pdf TOP
Joachim M. Buhmann, University of Bonn (Germany)
Thomas Hofmann, University of Bonn (Germany)
Competitive neural networks can be used to efficiently quantize image and video data. We discuss a novel class of vector quantizers which perform noise robust data compression. The vector quantizers are trained to simultaneously compensate channel noise and code vector elimination noise. The training algorithm to estimate code vectors is derived by the maximum entropy principle in the spirit of deterministic annealing. We demonstrate the performance of noise robust codebooks with compression results for a teleconferencing system on the basis of a wavelet image representation.
ic970139.pdf ic970139.pdf TOPThomas Vetter, Max-Planck-Institut, Tübingen (Germany)
A new technique is described for recognizing faces from new viewpoints. From a single 2D image of a face synthetic images from new viewpoints are generated and compared to stored views. A novel 2D image of a face can be computed without knowledge about the 3D structure of the head. The technique draws on prior knowledge of faces based on example images of other faces seen in different poses and on a single generic 3D model of a human head. The example images are used to learn a pose-invariant shape and texture description of a new face. The 3D model is used to solve the correspondence problem between images showing faces in different poses. The performance of the technique is tested on a date set of 200 faces of known orientation for rotations up to 90 degree.
ic970143.pdf ic970143.pdf TOP
Wolfgang Utschick, Technical University of Munich (Germany)
Josef A. Nossek, Technical University of Munich (Germany)
An extension of a feedforward neural network is presented. Although utilizing linear threshold functions and a boolean function in the second layer, signal processing within the neural network is real. After mapping input vectors onto a discretization of the input space, real valued features of the internal representation of pattern are extracted. A vectorquantizer assigns a class hypothesis to a pattern based on its extracted features and adequate reference vectors of all classes in the decision space of the output layer. Training consists of a combination of combinatorial and convex optimization. This work has been applied to a standard optical character recognition task. Results and comparison to alternative approaches are presented.
ic970147.pdf ic970147.pdf TOP
Yann LeCun, AT&T Labs (U.S.A.)
Léon Bottou, AT&T Labs (U.S.A.)
Yoshua Bengio, AT&T Labs (U.S.A.)
We propose a new machine learning paradigm called Multilayer Graph Transformer Network that extends the applicability of gradient-based learning algorithms to systems composed of modules that take graphs as input and produce graphs as output. A complete check reading system based on this concept is described. The system combines convolutional neural network character recognizers with graph-based stochastic models trained cooperatively at the document level. It is deployed commercially and reads million of business and personal checks per month with record accuracy.
ic970151.pdf ic970151.pdf TOP
Martin Schlang, Siemens AG, Munich (Germany)
Einar Broese, Siemens AG, Munich (Germany)
Bjoern Feldkeller, Siemens AG, Munich (Germany)
Otto Gramckow, Siemens AG, Munich (Germany)
Michael Jansen, Siemens AG, Munich (Germany)
Thomas Poppe, Siemens AG, Munich (Germany)
Clemens Schaeffner, Siemens AG, Munich (Germany)
Guenter Soergel, Siemens AG, Munich (Germany)
Neural Networks are particularly suitable for the approximation of non-linear time-variant functions. Due to their learning capabilities, they have proven useful in control applications for complex industrial processes. In collaboration with the Corporate Research and Development Department, the Siemens Industrial and Building Systems Group developed Neural Network applications for the steel industry, resulting in a more economic use of resources and an improvement of productivity. At this time Siemens has installed more than 100 neural nets world wide at different plants.
ic970155.pdf ic970155.pdf TOP
Peter Marbach, LIDS, MIT (U.S.A.)
John N. Tsitsiklis, LIDS, MIT (U.S.A.)
We are interested in solving large-scale Markov Decision Problems. The classical method of Dynamic Programming provides a mathematical framework for finding optimal solutions for a given Markov Decision Problem. However, for Dynamic Programming algorithms become computationally infeasible when the underlying Markov Decision Problem evolves over a large state space. In recent years, a new methodology, called Neuro-Dynamic Programming, has emerged which tries to overcome this ``curse of dimensionality''. We present how Neuro-Dynamic Programming can be applied to the Admission Control Problem for a single link in an ATM environment. Based on results obtained through Neuro-Dynamic Programming, we derive a heuristic ``Threshold'' policy. Performances of the policies obtained through Neuro-Dynamic Programming are compared with a policy which always accepts a customer when the required resources are available.
ic970159.pdf ic970159.pdf TOP