Information routing, correspondence finding, and object recognition in the brain [Elektronische Ressource] / von Philipp Wolfrum

Sujets

• Informatik

Informations

• Publié par: goethe_universitat_frankfurt_am_main
• Publié le: 01 janvier 2008
• Nombre de lectures: 11
• Langue: English
• Poids de l'ouvrage: 5 Mo

Extrait

Information Routing, Correspondence
Finding, and Object Recognition
in the Brain
DISSERTATION
zur
Erlangung des Grades
„Doktor der Naturwissenschaften“
vorgelegt beim Fachbereich Informatik und Mathematik
der Goethe-Universität Frankfurt am Main
von
Philipp Wolfrum
aus
Heilbronn
Frankfurt (2008)vom Fachbereich Informatik und Mathematik der
Goethe-Universität Frankfurt am Main als Dissertation angenommen.
Dekan: Prof. Dr. Klaus Johannson
1. Gutachter: Prof. Dr. Rudolf Mester
2. Gutachter: Prof. Dr. Christoph von der MalsburgAcknowledgments
This work would not have been possible without the people I have had the honor of working with
over the past years. First of all I want to thank Christoph von der Malsburg, whose scientific
concepts have inspired large parts of this thesis, and who granted me enough freedom to follow
my own research directions while at the same time providing motivation whenever this was
necessary. I also thank Rudolf Mester for good discussions that helped me retain my engineering
viewpoint on the interdisciplinary problems that I worked on. Discussions with Geoff Goodhill,
Bruno Olshausen, Jochen Triesch, and Alan Yuille, among others, have shaped my thinking and
had an important impact on this thesis.
I thank Urs Bergmann, Jenia Jitsev, and Junmei Zhu for proof-reading parts of the thesis, and
Jörg Lücke for good collaboration. The neurogroups at FIAS have provided a rich and stimu-
lating environment, it has been fun working and thinking with you! I also thank all colleagues
at FIAS for good company and the truly interdisciplinary interaction we had (e.g. at Hirschegg).
Last but not least I owe this thesis to my parents, who planted curiosity and a desire for under-
standing in me that were stronger than the difficulties I met during this dissertation.ivContents
List of Figures vii
List of Tables ix
1 Introduction 1
1.1 Object Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Strategies for Achieving Invariance . . . . . . . . . . . . . . . . . . . 2
1.1.2 State of the Art in Object Recognition . . . . . . . . . . . . . . . . . . 3
1.2 Plausibility of the Two Concepts . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.1 Computational Arguments . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.2 Experimental Evidence . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3 Proposal for Dynamic Routing . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2 Neural Model for Face Recognition 13
2.1 Correspondence Finding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Face Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3 The Basic Computational Units: Cortical Columns . . . . . . . . . . . . . . . 16
2.3.1 Neurobiological Background . . . . . . . . . . . . . . . . . . . . . . . 16
2.3.2 A Model of the Cortical Column . . . . . . . . . . . . . . . . . . . . . 18
2.4 The Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.4.1 Input Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.4.2 Assembly Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.4.3 Gallery Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.5.1 General Network Behavior . . . . . . . . . . . . . . . . . . . . . . . . 31
2.5.2 Position Invariance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.5.3 Tests on Standard Databases . . . . . . . . . . . . . . . . . . . . . . . 36
2.5.4 Attention Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3 Switchyards—Routing Structures in the Brain 47
3.1 Multi-Stage Routing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.2 Physiological Background of Dynamic Routing . . . . . . . . . . . . . . . . . 48
3.3 Optimized Architectures for Routing . . . . . . . . . . . . . . . . . . . . . . . 49
3.3.1 Routing Between Two Regions of the Same Size . . . . . . . . . . . . 50
3.3.2 Circuit with Different Sizes of Input and Output Layer . . . . 56vi Contents
3.4 Interpretation of Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.4.1 Difference to Sorting Networks . . . . . . . . . . . . . . . . . . . . . 58
3.4.2 Physiological Interpretation . . . . . . . . . . . . . . . . . . . . . . . 59
3.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4 Ontogenesis of Switchyards 67
4.1 Ontogenetic Plasticity Mechanisms in the Brain . . . . . . . . . . . . . . . . . 68
4.2 A Model for the Growth of Routing Networks . . . . . . . . . . . . . . . . . . 69
4.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.3.1 Noise Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.3.2 Growth of Three-Dimensional Networks . . . . . . . . . . . . . . . . 78
4.4 Other Potential Mechanisms . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
5 Recognition with Switchyards 85
5.1 Matching of Two Patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.2 Recognition from a Gallery of Patterns . . . . . . . . . . . . . . . . . . . . . . 93
5.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
6 Discussion and Outlook 101
Appendix 105
A Self-Normalization Properties of Columnar Dynamics . . . . . . . . . . . . . . 105
B Gabor Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
Bibliography 109
Zusammenfassung in deutscher Sprache 123
1 Einleitung . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
2 Neuronales Modell zur Gesichtserkennung . . . . . . . . . . . . . . . . . . . . 124
3 Switchyards . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
4 Ontogenese von Switchyards . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
5 Mustererkennung mit Switchyards . . . . . . . . . . . . . . . . . . . . . . . . 128
6 Diskussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
Lebenslauf 131List of Figures
1.1 Challenges for a vision system . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1 The correspondence problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 Columnar organization of cortex . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.3 Timecourse of unit activities . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.4 Principal layout of the face recognition system . . . . . . . . . . . . . . . . . . 22
2.5 Faces represented by a grid or a face graph . . . . . . . . . . . . . . . . . . . . 23
2.6 Architecture of the network . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.7 Information flow in the network . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.8 Average face graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.9 Interaction among control units . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.10 Matching process between Input and Input Assembly . . . . . . . . . . . . . . 32
2.11 Recognition process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.12 Position invariance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.13 A sample of 30 faces from the FERET database . . . . . . . . . . . . . . . . . 36
2.14 Cumulative match scores for the . . . . . . . . . . . . . . . . 37
2.15ve match for the AR database . . . . . . . . . . . . . . . . . . 38
2.16 Spatial attention experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.17 Object search e . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.18 Activity of the Gallery Assembly after priming of female faces . . . . . . . . . 43
2.19 Result of a priming experiment . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.1 One- and two-dimensional routing networks . . . . . . . . . . . . . . . . . . . 51
3.2 Number of required units as a function of intermediate layers . . . . . . . . . . 53
3.3 Prefactorsc andc~ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.4 Possible forms of tapered networks . . . . . . . . . . . . . . . . . . . . . . . . 56
3.5 Routing network with linear decrease of layer size . . . . . . . . . . . . . . . . 57
3.6 Number of possible conflicts as a function of distance of input nodes . . . . . . 60
3.7 Dependence of network size on parameter . . . . . . . . . . . . . . . . . . . 62
4.1 Axonal growth cone . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.2 Switchyard architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.3 TermG for alignment of coordinate systems . . . . . . . . . . . . . . . . . . . 72
4.4 Role of the marker similarity term . . . . . . . . . . . . . . . . . . . . . . . . 73
4.5 Snapshots of the growth process . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.6 Results for layer sizen = 27 . . . . . . . . . . . . . . . . . . . . . . . . . . . 76viii ListofFigures
4.7 Results for layer size 125 and n