Paper and thesis available: Spatial feature binding and segmentation

[Heiko Wersing]

Dear Connectionists,

I would like to announce the following recently accepted
paper and my PhD Thesis on spatial feature binding and learning in
competitive neural layer architectures.

The paper and thesis can be downloaded from my homepage at
http://www.techfak.uni-bielefeld.de/~hwersing/

Comments and questions are welcome.


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

"A competitive layer model for feature binding and sensory
segmentation."

by Heiko Wersing, Jochen J. Steil, and Helge Ritter
to appear in Neural Computation

Abstract:

We present a recurrent neural network for feature binding and
sensory segmentation, the competitive layer model (CLM). The CLM uses
topographically structured competitive and cooperative
interactions in a layered network to partition a set of input
features into salient groups. The dynamics is formulated within a
standard additive recurrent network with linear threshold neurons.
Contextual relations among features are coded by pairwise
compatibilities which define an energy function to be minimized by
the neural dynamics. Due to the usage of dynamical winner-take-all
circuits the model gains more flexible response properties than spin
models of segmentation by exploiting amplitude information in the
grouping process.  We prove analytic results on the convergence and
stable attractors of the CLM, which generalize earlier results on
winner-take-all networks, and incorporate deterministic annealing
for robustness against local minima.  The piecewise linear dynamics
of the CLM allows a linear eigensubspace analysis which we use to
analyze the dynamics of binding in conjunction with annealing.  For the
example of contour detection we show how the CLM can
integrate figure-ground segmentation and grouping into a unified
model.

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

"Spatial feature binding and learning in competitive neural
layer architectures"

by Heiko Wersing,
PhD Thesis, Faculty of Technology, University of Bielefeld, March 2000

Abstract:

The goal of this thesis is to contribute to the understanding of
feature binding processes by investigating an artificial neural
network model for spatial feature binding, the competitive layer model
(CLM). The CLM is a recurrent neural network, which employs
topographically structured competitive and cooperative interactions in
a system of neural layers to represent binding by the layer-wise
coactivation of feature-representing neurons. This model is explored
by means of mathematical analysis and simulations on artificial and
challenging real-world data.

The first chapter motivates the issue of feature binding as one of the
important questions regarding our understanding of brain function.
The second chapter gives an overview of the controversial scientific
discussion of the binding problem and reviews different neural network
model approaches to binding with a focus on their application in
sensory segmentation and perceptual grouping.

In the third chapter some new theoretical results on the stability of
general linear threshold networks are established, which allow to
operate the CLM in a mode of strong contextual modulation.  The
conditions provide a regime, where linear threshold networks are at
the same time sensitive to small changes in their inputs and are
capable of strong recurrent amplification without runaway activity.

In chapter four the competitive layer model is introduced and its
structure and dynamics are described. A deterministic annealing
mechanism is introduced, which is compared to Potts-spin mean field
models.  The stability conditions of chapter three are applied to
ensure convergence of the CLM, and additional conditions are proved,
which guarantee the Winner-Take-All behaviour of the competitive
columnar interactions. The grouping dynamics is characterized in
relation to the lateral interactions by performing an eigensubspace
analysis.  Finally, the lateral coupling scheme of the CLM is
generalized to general labeling problems and the chapter concludes
with a comparison to other labeling approaches.  Chapter five presents
the application of the CLM to model a wide range of Gestalt-based
perceptual grouping laws. First, the grouping of contours according to
the principle of continuity is considered by using oriented edge
elements as features. The results are discussed for real images, and
an extension of the contour grouping approach to the task of cell
segmentation is described.  Other grouping principles which are
considered include motion grouping, greyscale segmentation, and
texture segmentation.

In chapter six methods are presented to obtain appropriate lateral
interactions for grouping and segmentation by supervised learning
processes from manually labelled training patterns. The methods are
evaluated on an artificial data set and cell images from fluorescence
microscopy. A different learning approach is presented in Chapter
seven, which aims at optimizing the lateral connection
structure in order to improve the capabilities for solving complex
constraint satisfaction problems. The proposed backtracking
deterministic annealing method is interpreted as a heuristic approach
to combine classical backtracking with neural deterministic annealing
and successfully applied to the problem of complex tiling problems.
The concluding chapter summarizes the main results and
discusses possible future research directions.

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


-- 
Heiko Wersing
Future Technology Research
HONDA R&D EUROPE (DEUTSCHLAND) GmbH
Carl-Legien-Str. 30
63073 Offenbach /Main
Germany
Tel.: +49-69-89011741
Fax:  +49-69-89011749
e-mail: heiko.wersing at hre-ftr.f.rd.honda.co.jp