Sequence generation in dynamic field theory [Elektronische Ressource] / von Yulia Sandamirskaya

Sandamirskaya, Yulia

Sujets

• Physik

Informations

• Publié par: ruhr-universitat_bochum
• Publié le: 01 janvier 2010
• Nombre de lectures: 38
• Langue: English
• Poids de l'ouvrage: 14 Mo

Extrait

Sequence generation
in Dynamic Field Theory
DISSERTATION
zur
Erlangung des Grades
\Doktor der Naturwissenschaften"
an der Fakult at fur Physik und Astronomie
der Ruhr-Universit at Bochum
von
Yulia Sandamirskaya
aus
Minsk, Belarus
Bochum, 20101. Gutachter: Prof. Dr. Gregor Sch oner
2. Gutachter: Prof. Dr. Laurenz Wiskott
Datum der Disputation: 16.12.2010
iiAbstract
Cognitive processes are responsible for the complexity of human behavior
that is not exclusively controlled by sensory input but depends on the inner
state of the nervous system. In embodied cognition, the link between the
neural processes underlying cognition and the sensory-motor processes that
connect an organism or agent to the physical world is taken seriously. This
means, in particular, that the temporal continuity and graded nature of
perceptual, cognitive, and motor processes and of the neural dynamics sup-
porting these are taken into account. Because cognitive and neural processes
may evolve on di erent time-scales, a theoretical challenge is to understand
how fast neural processes may generate slow behavioral processes.
This problem is particularly apparent in sequence generation, a central prob-
lem of embodied cognition. A fast neural dynamics producing a sequence
of states must generate behaviors that unfold in the physical world. In
the face of the inherent variability of sensory information, stability of the
behaviorally relevant neural states is a necessary property of any neural
sequence generation mechanism. On the other hand, stability is in con-
ict with sequence generation as the transition from one action to the next
requires that the previous neural state decays and a new neural state is
activated. In this thesis, I propose a mechanism that solves this stabil-
ity vs. sequentiality trade-o based on Dynamic Field Theory (DFT). The
DFT is a framework for understanding embodied cognition within which
models of cognitive functions can be formulated that enable coupling of the
neural dynamics supporting cognition to the sensory-motor systems sup-
porting behavior. The sequence generating model I introduce includes a
neural dynamic representation of a condition of satisfaction that signals the
accomplishment of an action and triggers the transition to the next action.In my model, actions are stable states of the neural eld dynamics. First, I
present a conceptual model in which a single metric dimension characterizes
actions in a sequence. That this model ful lls the constraints of embodied
cognition is demonstrated by implementing the model on a low-level robotic
vehicle that performs a color-search task. Autonomous sequence generation
is also demonstrated in an application to modeling turn taking in human
communication. Next, I introduce a modi ed architecture, in which the se-
rial order of actions, i.e. their ordinal position within a sequence, is encoded
by a discrete dynamics of bistable interconnected nodes that project their
activity onto distributed dynamic neural eld representations of actions.
This simpli cation makes it possible to analyze the ordinal dynamics. The
function of this model is demonstrated in the simple color-search task and
in a more complex robotic scenario that includes arm movement along with
the color-search behavior. Finally, I describe an application of this model
to control of the dynamics of a DFT model of spatial language behaviors.
The ordinal dynamics of my model enables autonomous processing within
this complex architecture. The proposed sequence generation model o ers
a principled solution to the problem of stability vs. sequentiality trade-o .
The model has been implemented and tested in several settings and may
constitute the rst step toward the development of a complete architecture
that controls autonomous behavioral sequences of embodied and situated
agents. Analysis of the constraints of embodiment on sequence generation
advance our understanding also of more abstract forms of cognitive process-
ing.To my family and friendsAcknowledgements
Many thanks to everyone who supported me during my doctorate years in Bochum
and who helped in one way or another in the work that has led to this thesis.
Thanks to all colleagues at the Institut fur Neuroinformatik, I really enjoyed the
friendly and collaborative atmosphere there. Special and personal thanks to: Chris-
tian Faubel for supporting me in learning c++ and for sharing his inspiration and
excitement about our research, a dense exchange of ideas with him on an early
stage of my work helped me to have a fast start, the challenging discussions stim-
ulated thinking and furious arguments sharpened understanding; Sebastian Noth
for the numerous hints and tricks that helped me to design the software for my
research, for his readiness to help in all times; John Lipinski for proof reading, for
discussions, and also for a brilliant and very fruitful collaboration on the spatial
language project; Jannis Iossi dis for many advices that helped to organize infras-
tructure around o ce and lab; Sebastian Schneegans for many conversations and
discussions, his readiness to listen and his weighed and reasonable views, and also
for the collaboration; Alex Gepperth for a fun time during the deep-drill project
and for many project ideas (I’m sorry that I didn’t nd time to realize most of
them). I would like to thank my latest o ce mate Hendrik Reimann for a friendly
working mood in the o ce, Urun Dogan and Andrey Bogdanov for the many infor-
mative corridor-chats. Many thanks to Prof. Dr. Laurenz Wiskott for a thorough
review of the dissertation, many useful comments, suggestions for improvements
and future work. Also, I’m very specially thankful to my parents and my brother,
whose support during these years made this step in my scienti c career and life
possible. I thank also my kids for their patience with the busy Mom. And last,
but not the least, I would like to thank my scienti c supervisor, Prof. Dr. Gregor
Sch oner, whose ideas inspired and motivated me during these years, whose views
shaped my scienti c thinking, whose critics improved my writing and presenting
style, and whose scienti c brilliance I keep admiring.Contents
List of Figures vii
List of Tables xi
1 Introduction 1
1.1 Preamble . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Motivation for the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.1 Motivation within dynamical systems approach to cognition . . . 2
1.2.2 Motivation within the problem of serial order. . . . . . . . . . . . 4
1.3 Main goals of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4 Publications of the results of the thesis . . . . . . . . . . . . . . . . . . . 7
1.5 Structure of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2 Background and methods 9
2.1 State of the art in modeling sequence generation . . . . . . . . . . . . . 9
2.1.1 Behavioral signatures of serial order . . . . . . . . . . . . . . . . 10
2.1.2 Neurophysiology of serial order . . . . . . . . . . . . . . . . . . . 11
2.1.3 Classi cation of models of serial order according to Henson . . . 12
2.1.4 Connectionist models . . . . . . . . . . . . . . . . . . . . . . . . 13
2.1.5 Neural dynamic models . . . . . . . . . . . . . . . . . . . . . . . 14
2.1.6 Sequence generation in robotics . . . . . . . . . . . . . . . . . . . 21
2.1.7 Open issues in modeling sequential action and the problem state-
ment for the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.2 Dynamic Field Theory to cognition . . . . . . . . . . . . . . . . . . . . . 22
2.2.1 Dynamic equation of a neural eld . . . . . . . . . . . . . . . . . 22
2.2.2 regimes and instabilities of a dynamic neural eld . . . 23
iiiCONTENTS
2.2.3 Dynamics of discrete neural nodes. . . . . . . . . . . . . . . . . . 25
2.2.4 Coupling between neural elds and nodes . . . . . . . . . . . . . 26
2.2.5 Problem of sequentiality in DFT to cognition . . . . . . . . . . . 27
3 The DFT sequence generation architecture. A one-dimensional model. 29
3.1 The task setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.2 The DFT architecture for sequence generation . . . . . . . . . . . . . . . 30
3.2.1 Ordinal elds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.2.2 Sequence memory: preshape of the ordinal elds . . . . . . . . . 32
3.2.3 Output eld . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.2.4 Condition of satisfaction system . . . . . . . . . . . . . . . . . . 33
3.3 Mathematical description of the model . . . . . . . . . . . . . . . . . . . 35
3.4 Robotic implementation of the DFT sequence generation architecture . 36
3.5 Results of robotic demonstrations . . . . . . . . . . . . . . . . . . . . . . 40
3.5.1 Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.5.2 Sequence generation . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4 Application of the DFT sequence generation mechanism to model
turn taking 45
4.1 Embodied communication and turn taking: motivation for this application 45
4.2 The DFT model of turn taking . . . . . . . . . . . . . . . . . . . . . . . 46
4.3 Mathemati