2nd and revised edition
Keywords: Biological Cybernetics, System Theory, Artificial Neural Networks, Simulation, Tutorial
Biological systems are usually much too complicated to be understood in their entirety. Scientific progress is therefore generally based on the fragmentation of the systems under investigation. This means the system is broken down into smaller parts or subsystems, which can then be more easily approached. However, such a reduction to a lower level - for instance - from behavior to reflexes, from reflexes to neuronal cells, or from cells to molecules - also has serious shortcomings. First, the overview of the whole system may be lost. Looking at the lower level, one may not see "the forest for the trees" because many system properties are only understandable when not only the individual parts of the system, but also the cooperation of these parts are taken into account. Second, this reduction may further increase the gap that already exists between biological research and the investigation of problems on more complex levels, such as those considered in psychology or even philosophy. To make this gap smaller "holistic" approaches are required.
One sensible way to oppose this reductionistic path is the use of simulation. The construction of quantitative models, usually in the form of computer simulation, is an important tool in biology. Such a simulation allows a step in the other, "antireductionistic" direction, namely to construct complex systems from smaller, simple elements. Through investigation of the simulated system, the properties of the whole system can be better understood.
The tool of simulation is particularly important for understanding the dynamic properties of systems. Such dynamic properties are often produced by feedback loops within the system. However, the human brain does not seem very well adapted to grasp such systems. Simulations could improve this situation. We might become more familiar with the properties of dynamic systems and thus train ourselves to understand such systems so that, even without an explicit computer simulation, some predictions could be made. Such dynamic systems occur in many fields, from genetics, metabolism, and ecology to, of course, neurobiology and ethology. Although this book will concentrate on the latter two, the tools provided can also be applied to the other fields. But these tools are also applicable to fields outside biology, e. g., psychology, and to even more distant areas, such as economics, physics, and electrotechnology (which in fact gave rise to many of these ideas).
Ethology, although an important field in biology, had been attracting less interest in recent decades mainly because of a lack of theoretical concepts. This has, however, dramatically changed in recent years because the emergence of the theory of artificial neural networks (ANN) and the field of artificial life, has led to the development of a great number of models and modeling tools that can now be fruitfully applied to ethology and neuroethology. Although the treatment of simple neuronal models was already an important subject of early biological cybernetics, the focus of interest later moved to "pure" systems theory. Only in the last decade did the field of ANN approach gain its enormous thrust. These two fields have not yet been brought into intensive contact with each other, but the consideration of dynamic properties so central to systems theory has the potential to make a great impact on ANN. This book attempts to combine both approaches which, as mentioned, stem from the same roots. It can be expected that both fields will profit from each other.
Usually textbooks on these fields are loaded down with a great deal of mathematics, which makes them somewhat indigestible for the typical student of biology. To minimize these problems, this book tries to avoid the use of mathematical formulas as far as possible. The text is based not on differential equations or on complex variables, but rather on illustrations. It nevertheless provides sufficient information to permit the reader to develop quantitative models. Technical aspects are relegated to the appendices.
The first part of this book is based on an earlier version of my book "Biologische Kybernetik", which was inspired by the exciting lectures of D. Varju at the University of Tübingen. Writing this greatly expanded version of the book would not have been possible without the help of a number of colleagues who met in a research group funded by the Center of Interdisciplinary Research (ZiF) of the University of Bielefeld. I mainly want to thank to H. U. Bauer, Frankfurt; H. Braun, Karlsruhe; G. Hartmann, Paderborn; J. Dean, A. Dress, P Lanz, H. Ritter and J. Schmitz, all from Bielefeld; and H. Scharstein, Cologne, for a number of helpful comments. Furthermore, l would like to thank to A. Baker who helped with the English in an earlier version of the manuscript, A. Exter for the preparation of many figures, and P. Sirocka and Th. Kindermann for providing Figures B 5.3 and B 5.1, respectively. Furthermore, I owe a debt of gratitude to many students of my lectures who succeeded in finding errors and unclear formulations in the text. Of course, the responsibility for all remaining flaws is my own.
April 1996 Holk Cruse
Ten years ago, the first edition of the book ”Neural Networks as Cybernetics Systems“ has been published by Thieme. At this time there was still an ongoing debate whether the neural network approach comprises just a fashionable, but short living hype. Meanwhile this approach is well established. Understanding complex systems by means of simulation is more and more accepted, also within biology. The more it is important to provide students with a tool that helps to understand and also actively perform simulations. This is particularly important for students their primary education was not in mathematics or in computer science. It is the goal of this text to provide such a tool. In the first part, both linear and nonlinear aspects of systems theory, sometimes called filter theory or theory of dynamical systems, are treated in a way that mathematical terms are avoided as far as possible. In part II this is extended to the theory of massively parallel systems, or theory of neural networks. This qualitative approach is also suited as a first step for students that later plan to follow a more thorough quantitative understanding.
Appearing as an (open access) e-version, the handling of the text is easier compared to the earlier printed version, figures are coloured, errors are corrected, (hopefully new errors appear at a minimum), and some chapters, in particular those concerning the important field of recurrent networks, are added anew. The most helpful extension however concerns the software package that allows to perform exercises concerning simulations for part I. This package is written so that it can be used in an extremely simple way. Extensions for part II are in preparation.
January 2006 Holk Cruse
Any party may pass on this Work by electronic means and make it available for download under the terms and conditions of the Digital Peer Publishing License. The text of the license may be accessed and retrieved via Internet at http://www.dipp.nrw.de/lizenzen/dppl/dppl/DPPL_v2_en_06-2004.html