Thesis Announcement

[Richard Woesler]

--------------------------------------------------------------------------------
Dear Connectionists,

I am pleased to announce the availability of my thesis.

Regards,
Richard Woesler

--------------------------------------------------------------------------------

Title: 	Neural Networks for Object Segmentation in Amphibians 
           and an algorithmic Classification of Neurons

ISBN 3-89722-073-3, Logos-Verlag Berlin, Germany 1999

Advisors: 
Prof. Helmut Schwegler, head of the Institute of Theoretical Neurophysics at the 
University of Bremen, and Prof. Gerhard Roth, head of the Institute of Brain 
Research at the University of Bremen, Germany 

Language: german

Abstract:
Neural network models for object segmentation in amphibians, especially in the 
tongue projecting salamander Hydromantes italicus, constitute the main part of 
the text. Firstly, we carry out a simple double-dummy experiment in which the 
animal has to choose one dummy out of two identical, coherently moving 
prey-dummies. Hydromantes is able to select one of them and to snap at it with 
high accuracy. The question arises how this object-selection process, the 
segmentation of the selected object from the rest of the visual scene, and the 
high accuracy are realized within the amphibian brain. 
First, we present a simplified model, our Base Model. This model detects moving 
contrast edges, binds these edges to objects, and selects one object. We derive 
analytical results concerning the performance level of the network. The results 
are similar to those known from amphibians. 
Second, we extend this model to our Amphibian Model where a number of recent 
biological data is taken into account, especially data about the retina and the 
tectum opticum which is an important brain region with respect to prey-capture 
behaviour and which is homologous to the mammalian superior colliculus, and data 
about the nucleus isthmi which is a relatively small brain region connected 
reciprocally with the tectum opticum and which is homologous to the mammalian 
nucleus parabigeminalis. This model is a complete neural network for object 
segmentation from retinal photoreceptors to the segmentation of the selected 
object. 
The model yields possible explanations for many biological findings: e. g., for 
the width distribution of the receptive fields of neurons of the tectum opticum, 
or for the range of velocities of objects for which Hydromantes italicus shows 
prey-capture behaviour.
The model contains a spotlight network which enhances the retinotectal 
transmission in a certain spotlight region to further neurons, which we call the
multiplicative coarse coding neurons. The spotlight network should segment the 
selected object from the rest of the visual scene, and the multiplicative coarse 
coding neurons should encode its place with high accuracy. We describe in 
detail how this network can be realized in amphibians.
Third, we present the computer model Simulander III with which we simulate the 
described double-dummy experiment. Indeed the model shows the high 
object-selection ability and the high accuracy known from Hydromantes. We give 
an outlook to various model extensions, especially to a model of recognition of 
the absolute width of the selected object and of recognition of a stepwise prey 
movement. These abilities are also known from Hydromantes. 
Additionally, we discuss a model extension which includes a neural network of 
synchronizing neurons.

In the second part of the text, we consider the classification of neurons of the 
tectum opticum of Hydromantes italicus according to their responses to different 
prey-dummies. We are particularly interested in the question whether the 
response types form distinct classes or a continuum. We classify the data with 
an algorithm that we have developed using known classification methods. The 
classification result suggests that the data form a continuum with some 
accumulations. Finally, we discuss the possible biological reasons for a 
continuum and possible advantages of a continuum.
--------------------------------------------------------------------------------


********************************************
Dipl.-Phys. Dr. Richard Woesler
August-Schlueter-Str. 39
48249 Duelmen
Germany

Phone +49 (0) 2594-991182
Fax   +49 (0) 2594-948752
http://home.t-online.de/home/Richard.Woesler
********************************************