Biologically Inspired Cognitive Architectures

2010 Proceedings Of The First Annual Meeting Of

The Bica Society 1st Edition A V Samsonovich K R
Jhannsdttir A Chella download
https://ebookbell.com/product/biologically-inspired-cognitive-
architectures-2010-proceedings-of-the-first-annual-meeting-of-
the-bica-society-1st-edition-a-v-samsonovich-k-r-jhannsdttir-a-
chella-51692978

Explore and download more ebooks at ebookbell.com

Here are some recommended products that we believe you will be
interested in. You can click the link to download.

Biologically Inspired Cognitive Architectures Bica For Young

Scientists Proceedings Of The First International Early Research
Career Enhancement School Fierces 2016 1st Edition Alexei V
Samsonovich
https://ebookbell.com/product/biologically-inspired-cognitive-
architectures-bica-for-young-scientists-proceedings-of-the-first-
international-early-research-career-enhancement-school-
fierces-2016-1st-edition-alexei-v-samsonovich-5484574

Biologically Inspired Cognitive Architectures Bica For Young

Scientists Proceedings Of The First International Early Research
Career Enhancement School On Bica And Cybersecurity Fierces 2017
Klimov
https://ebookbell.com/product/biologically-inspired-cognitive-
architectures-bica-for-young-scientists-proceedings-of-the-first-
international-early-research-career-enhancement-school-on-bica-and-
cybersecurity-fierces-2017-klimov-6753320

Biologically Inspired Cognitive Architectures 2012 Proceedings Of The

Third Annual Meeting Of The Bica Society 1st Edition Igor Aleksander

https://ebookbell.com/product/biologically-inspired-cognitive-
architectures-2012-proceedings-of-the-third-annual-meeting-of-the-
bica-society-1st-edition-igor-aleksander-4230686

Biologically Inspired Cognitive Architectures 2019 Proceedings Of The

Tenth Annual Meeting Of The Bica Society 1st Ed Alexei V Samsonovich

https://ebookbell.com/product/biologically-inspired-cognitive-
architectures-2019-proceedings-of-the-tenth-annual-meeting-of-the-
bica-society-1st-ed-alexei-v-samsonovich-10489210
Biologically Inspired Cognitive Architectures 2018 Proceedings Of The
Ninth Annual Meeting Of The Bica Society 1st Ed Alexei V Samsonovich

https://ebookbell.com/product/biologically-inspired-cognitive-
architectures-2018-proceedings-of-the-ninth-annual-meeting-of-the-
bica-society-1st-ed-alexei-v-samsonovich-7325490

Biologically Inspired Cognitive Architectures 2024 Proceedings Of The

15th Annual Meeting Of The Bica Society 1st Edition Alexei V
Samsonovich

https://ebookbell.com/product/biologically-inspired-cognitive-
architectures-2024-proceedings-of-the-15th-annual-meeting-of-the-bica-
society-1st-edition-alexei-v-samsonovich-189182742

Biologically Inspired Cooperative Computing Ifip 19th World Computer

Congress Tc 10 1st Ifip International Conference On Biologically
Inspired Cooperative In Information And Communication Technology 1st
Edition Yi Pan
https://ebookbell.com/product/biologically-inspired-cooperative-
computing-ifip-19th-world-computer-congress-tc-10-1st-ifip-
international-conference-on-biologically-inspired-cooperative-in-
information-and-communication-technology-1st-edition-yi-pan-2141834

Biologically Inspired Approaches To Advanced Information Technology

First International Workshop Bioadit 2004 Lausanne Switzerland January
2930 2004 Revised Selected Papers 1st Edition Jacqueline Signorini

https://ebookbell.com/product/biologically-inspired-approaches-to-
advanced-information-technology-first-international-workshop-
bioadit-2004-lausanne-switzerland-january-2930-2004-revised-selected-
papers-1st-edition-jacqueline-signorini-2330358

Biologically Inspired Approaches For Locomotion Anomaly Detection And

Reconfiguration For Walking Robots 1st Edition Bojan Jakimovski Auth

https://ebookbell.com/product/biologically-inspired-approaches-for-
locomotion-anomaly-detection-and-reconfiguration-for-walking-
robots-1st-edition-bojan-jakimovski-auth-2451800
 BIOLOGICALLY INSPIRED COGNITIVE
ARCHITECTURES 2010
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 Frontiers in Artificial Intelligence and
Applications
FAIA covers all aspects of theoretical and applied artificial intelligence research in the form of
monographs, doctoral dissertations, textbooks, handbooks and proceedings volumes. The FAIA
series contains several sub-series, including “Information Modelling and Knowledge Bases” and
“Knowledge-Based Intelligent Engineering Systems”. It also includes the biennial ECAI, the
European Conference on Artificial Intelligence, proceedings volumes, and other ECCAI – the
European Coordinating Committee on Artificial Intelligence – sponsored publications. An
editorial panel of internationally well-known scholars is appointed to provide a high quality
selection.

Series Editors:
J. Breuker, N. Guarino, J.N. Kok, J. Liu, R. López de Mántaras,
R. Mizoguchi, M. Musen, S.K. Pal and N. Zhong

Volume 221
Recently published in this series

Vol. 220. R. Alquézar, A. Moreno and J. Aguilar (Eds.), Artificial Intelligence Research and
Development – Proceedings of the 13th International Conference of the Catalan
Association for Artificial Intelligence
Vol. 219. I. Skadiņa and A. Vasiļjevs (Eds.), Human Language Technologies – The Baltic
Perspective – Proceedings of the Fourth Conference Baltic HLT 2010
Vol. 218. C. Soares and R. Ghani (Eds.), Data Mining for Business Applications
Vol. 217. H. Fujita (Ed.), New Trends in Software Methodologies, Tools and Techniques –
Proceedings of the 9th SoMeT_10
Vol. 216. P. Baroni, F. Cerutti, M. Giacomin and G.R. Simari (Eds.), Computational Models of
Argument – Proceedings of COMMA 2010
Vol. 215. H. Coelho, R. Studer and M. Wooldridge (Eds.), ECAI 2010 – 19th European
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

Conference on Artificial Intelligence

Vol. 214. I.-O. Stathopoulou and G.A. Tsihrintzis, Visual Affect Recognition
Vol. 213. L. Obrst, T. Janssen and W. Ceusters (Eds.), Ontologies and Semantic Technologies
for Intelligence
Vol. 212. A. Respício et al. (Eds.), Bridging the Socio-Technical Gap in Decision Support
Systems – Challenges for the Next Decade
Vol. 211. J.I. da Silva Filho, G. Lambert-Torres and J.M. Abe, Uncertainty Treatment Using
Paraconsistent Logic – Introducing Paraconsistent Artificial Neural Networks
Vol. 210. O. Kutz et al. (Eds.), Modular Ontologies – Proceedings of the Fourth International
Workshop (WoMO 2010)

ISSN 0922-6389 (print)

ISSN 1879-8314 (online)

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 Biologically Inspired Cognitive
Architectures 2010
Proceedings of the First Annual Meeting of the BICA Society

Edited by
Alexei V. Samsonovich
George Mason University, USA

Kamilla R. Jóhannsdóttir
University of Akureyri, Iceland

Antonio Chella
University of Palermo, Italy
and
Ben Goertzel
Novamente LLC, USA
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

Amsterdam • Berlin • Tokyo • Washington, DC

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 © 2010 The authors and IOS Press.

All rights reserved. No part of this book may be reproduced, stored in a retrieval system,
or transmitted, in any form or by any means, without prior written permission from the publisher.

ISBN 978-1-60750-660-7 (print)

ISBN 978-1-60750-661-4 (online)
Library of Congress Control Number: 2010938639

Publisher
IOS Press BV
Nieuwe Hemweg 6B
1013 BG Amsterdam
Netherlands
fax: +31 20 687 0019
e-mail: [email protected]

Distributor in the USA and Canada

IOS Press, Inc.
4502 Rachael Manor Drive
Fairfax, VA 22032
USA
fax: +1 703 323 3668
e-mail: [email protected]
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

LEGAL NOTICE
The publisher is not responsible for the use which might be made of the following information.

PRINTED IN THE NETHERLANDS

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 Biologically Inspired Cognitive Architectures 2010 v
A.V. Samsonovich et al. (Eds.)
IOS Press, 2010
© 2010 The authors and IOS Press. All rights reserved.

Preface
This volume documents the proceedings of the First International Conference on Bio-
logically Inspired Cognitive Architectures (BICA 2010), which is also the First Annual
Meeting of the BICA Society. This conference was preceded by 2008 and 2009 AAAI
Fall Symposia on BICA that were similar in content (indeed, the special issue of the
International Journal of Machine Consciousness1 is composed of a selection of papers
and abstracts from all three events, and it is an official complement of this Proceedings
volume). The 2008–2009 BICA symposia in turn were preceded by a sequence of the
DARPA BICA meetings in 2005–2006 (see below). However, BICA 2010 is the first
independent event in the BICA series: it has the status of the first annual meeting of the
just established BICA society (further information is available at http://bicasociety.org).
Like the 2008 and 2009 BICA Symposia, the present BICA 2010 conference con-
tains a wide variety of ideas and approaches, all centered around the theme of under-
standing how to create general-purpose humanlike artificial intelligence using inspira-
tions from studies of the brain and the mind. BICA is no modest pursuit: the long-term
goals are no less than understanding how human and animal brains work, and creating
artificial intelligences with comparable or greater functionality. But, in addition to
these long-term goals, BICA research is also yielding interesting and practical research
results right now.
A cognitive architecture, broadly speaking, is a computational framework for the
design of intelligent and even conscious agents. Cognitive architectures may draw their
inspiration from many sources, including pure mathematics or physics or abstract theo-
ries of cognition. A biologically inspired cognitive architecture (BICA), in particular, is
one that incorporates formal mechanisms from computational models of human and
animal cognition, drawn from cognitive science or neuroscience. The appeal of the
BICA approach should be obvious: currently human and animal brains provide the only
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

physical examples of the level of robustness, flexibility, scalability and consciousness

that we want to achieve in artificial intelligence. So it makes sense to learn from them
regarding cognitive architectures: both for research aimed at closely replicating human
or animal intelligence, and also for research aimed at creating and using human-level
artificial intelligence more broadly.
Research on the BICA approach to intelligent agents has focused on several differ-
ent goals. Some BICA projects have a primary goal of accurately simulating human
behavior, either for purely scientific reasons – to understand how the human mind
works – or for applications in domains such as entertainment, education, military train-
ing, and the like. Others are concerned with even deeper correspondence between mod-
els and the human brain, going down to neuronal and sub-neuronal level. The goal in
this approach is to understand how the brain works. Yet another approach is concerned
with designing artificial systems that are successful, efficient, and robust at performing

1
A.V. Samsonovich (guest editor): International Journal of Machine Consciousness, special issue on Bio-
logically Inspired Cognitive Architectures, Vol. 2, No. 2, 2010.

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 vi

cognitive tasks that today only humans can perform, tasks that are important for practi-
cal applications in the human society and require interaction with humans. Finally,
there are BICA projects aimed broadly at creating generally intelligent software sys-
tems, without focus on any one application area, but also without a goal of closely
simulating human behavior. All four of these goals are represented in the various pa-
pers contained in this volume.
The term BICA was coined in 2005 by Defense Advanced Research Projects
Agency (DARPA), when it was used as the name of a DARPA program administered
by the Information Processing Technology Office (IPTO). The DARPA BICA program
was terminated in 2006 (more details are available at the DARPA BICA web page at
http://www.darpa.mil/ipto/programs/bica/bica.asp). Our usage of the term “BICA” is
similar to its usage in the DARPA program; however, the specific ideas and theoretical
paradigms presented in the papers here include many directions not encompassed by
DARPA’s vision at that time. Moreover, there is no connection between DARPA and
the BICA Society.
One of the more notable aspects of the BICA approach is its cross-disciplinary na-
ture. The human mind and brain are not architected based on the disciplinary bounda-
ries of modern science, and to understand them almost surely requires rich integration
of ideas from multiple fields including computer science, biology, psychology and
mathematics. The papers in this volume reflect this cross-disciplinarity in many ways.
Another notable aspect of BICA is its integrative nature. A well-thought BICA has
a certain holistic integrity to it, but also generally contains multiple subsystems, which
may in some cases be incorporated into different BICAs, or used in different ways than
the subsystem’s creator envisioned. Thus, the reader who is developing their own ap-
proach to cognitive architectures may find many insights in the papers contained here
useful for inspiring their own work or even importing into their own architecture, di-
rectly or in modified form.
Finally we would like to call attention to the relationship between cognition, em-
bodiment and development. In our view, to create a BICA with human-level general
intelligence, it may not be necessary to engineer all the relevant subsystems in their
mature and complete form. Rather, it may be sufficient to understand the mechanisms
of cognitive growth in a relatively simple form, and then let the mature forms arise via
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

an adaptive developmental process. In this approach, one key goal of BICA research
becomes understanding what the key cognitive subsystems are, how do they develop,
and how they become adaptively integrated in a physical or virtual situated agent able
to perform tight interactions within its own body, the other entities and the surrounding
environment. With this sort of understanding in hand, it might well be possible to cre-
ate a BICA with human-level general learning capability, and teach it like a child. Po-
tentially, a population of such learners could ignite a cognitive chain reaction of learn-
ing from each other and from common resources, such as human instructors or the
Internet.
Currently BICA research is still at an early stage, and the practical capabilities of
BICA systems are relatively undeveloped. Furthermore, the relationships between the
ideas of various researchers in the field are not always clear; and there is considerable
knowledge in relevant disciplines that is not yet fully incorporated into our concrete
BICA designs. But BICA research is also rapidly developing, with each year bringing
significant new insights, moving us closer to our ambitious goals. In this sense, the
newborn BICA society, according to the intentions of the Founding Members, will be a
main vehicle for the growth and dissemination of breakthrough research in the field of

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 vii

BICA systems. The papers presented in this volume form part of this ongoing process,
as will the papers in the ongoing BICA conferences to follow.

Alexei V. Samsonovich, Kamilla R. Jóhannsdóttir,

Antonio Chella and Ben Goertzel
Editors
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 This page intentionally left blank
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 ix

BICA 2010 Conference Committees

Organizing Committee
Chairs
Alexei V. SAMSONOVICH
George Mason University, USA
Kamilla R. JÓHANNSDÓTTIR
University of Akureyri, Iceland

Core
Igor Aleksander (Imperial College London, UK)
Bernard J. Baars (The neurosciences Institute, USA)
Antonio Chella (University of Palermo, Italy)
Ben Goertzel (Novamente LLC, USA)
Stephen Grossberg (Boston University, USA)
Christian Lebiere (Carnegie Mellon University, USA)
David C. Noelle (University of California Merced, USA)
Roberto Pirrone (University of Palermo, Italy)
Frank E. Ritter (Penn State University, USA)
Murray P. Shanahan (Imperial College London, UK)
Kristinn R. Thorisson (CADIA; Reykjavik University, Iceland)
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

Program Committee
Samuel S. Adams (Watson IBM Research, USA)
Itamar Arel (University of Tennessee, USA)
Son K. Dao (HRL Laboratories, LLC, USA)
Scott E. Fahlman (Carnegie Mellon University, USA)
Ian Fasel (University of Arizona, USA)
Stan Franklin (University of Memphis, USA)
Eva Hudlicka (Psychometrix Assoc., USA)
Magnus Johnsson (Lund University Cognitive Science, Sweden)
Alexander A. Letichevsky (Glushkov Institute of Cybernetics, Ukraine)
Ali A. Minai (University of Cincinnati, USA)
Shane T. Mueller (Klein Associates Division / ARA Inc., USA)
Brandon Rohrer (Sandia National Laboratories, USA)
Ricardo Sanz (Universidad Politécnica de Madrid, Spain)
Colin T. Schmidt (Le Mans University & Arts et Metiers ParisTech, France)
Josefina Sierra (Technical University of Catalonia, Spain)

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 x

Terry Stewart (University of Waterloo, Canada)

Andrea Stocco (Carnegie Mellon University, USA)
Bruce Swett (Decisive Analytics Corporation, USA)
Rodrigo Ventura (Instituto Superior Técnico, Portugal)
Pei Wang (Temple University, USA)
Juyang (John) Weng (Michigan State University, USA)

Reviewers
James S. Albus Roberto Pirrone
Igor Aleksander Lorenzo Riano
Itamar Arel Frank E. Ritter
Bernard J. Baars Brandon Rohrer
Jonathan Brickliln Paul Rosenbloom
Antonio Chella Alexei Samsonovich
Son K. Dao Ricardo Sanz
Scott E. Fahlman Colin T. Schmidt
Ian Fasel Michael Sellers
Stan Franklin Murray P. Shanahan
Ben Goertzel Josefina Sierra
Stephen Grossberg Terry Stewart
Wan Ching Ho Andrea Stocco
Eva Hudlicka Leopold Stubenberg
Kamilla R. Jóhannsdóttir Bruce Swett
Magnus Johnsson Kristinn R. Thórisson
Benjamin Johnston Peter Tripodes
Christian Lebiere Akshay Vashist
Alexander A. Letichevsky Robert N. VanGulick
Ali A. Minai Craig M. Vineyard
Jonathan H. Morgan Rodrigo Ventura
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

Shane T. Mueller Pei Wang

David C. Noelle Mark Waser
Rony Novianto Juyang (John) Weng

BICA 2010 conference was held Friday, Saturday and Sunday, November 12–14, 2010,
in Arlington, Virginia, USA.

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 xi

Contents
Preface v
Alexei V. Samsonovich, Kamilla R. Jóhannsdóttir, Antonio Chella and
Ben Goertzel
BICA 2010 Conference Committees ix

Conference Papers and Extended Abstracts

Reverse Engineering the Vision System 3

James Albus
Application Feedback in Guiding a Deep-Layered Perception Model 4
Itamar Arel and Shay Berant
NeuroNavigator: A Biologically Inspired Universal Cognitive Microcircuit 10
Giorgio A. Ascoli and Alexei V. Samsonovich
BINAReE: Bayesian Integrated Neural Architecture for Reasoning and
Explanation 17
Robert (Rusty) Bobrow, Paul Robertson and Robert Laddaga
SCA-Net: A Sensation-Cognition-Action Network for Speech Processing 23
Michael Connolly Brady
A Connectionist Model of MT+/Mstd Explains Human Heading Perception in
the Presence of Moving Objects 24
N. Andrew Browning
Discovering the Visual Patterns Elicited by Human Scan-Path 25
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

Andrea Carbone
An Architecture for Humanoid Robot Expressing Emotions and Personality 33
Antonio Chella, Rosario Sorbello, Giorgio Vassallo and Giovanni Pilato
An Evolutionary Approach to Building Artificial Minds 40
James L. Eilbert
Explanatory Aspirations and the Scandal of Cognitive Neuroscience 42
Ross W. Gayler, Simon D. Levy and Rens Bod
An Experimental Cognitive Robot 52
Pentti O.A. Haikonen
Dopamine and Self-Directed Learning 58
Seth Herd, Brian Mingus and Randall O’Reilly
Modelling Human Memory in Robotic Companions for Personalisation and
Long-Term Adaptation in HRI 64
Wan Ching Ho, Kerstin Dautenhahn, Mei Yii Lim and Kyron Du Casse

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 xii

Assessing the Role of Metacognition in GMU BICA 72

Michael Q. Kalish, Alexei V. Samsonovich, Mark A. Coletti and
Kenneth A. De Jong
Towards Understanding Trust Through Computational Cognitive Modeling 78
William G. Kennedy
An Externalist and Fringe Inspired Cognitive Architecture 79
Riccardo Manzotti
Architecture of the Mind with Artificial Neurons 85
Deepak J. Nath
Online Event Segmentation in Active Perception Using Adaptive Strong
Anticipation 86
Bruno Nery and Rodrigo Ventura
Four Kinds of Learning in One Agent-Oriented Environment 92
Sergei Nirenburg, Marjorie McShane, Stephen Beale, Jesse English
and Roberta Catizone
Attention in the ASMO Cognitive Architecture 98
Rony Novianto, Benjamin Johnston and Mary-Anne Williams
On the Emergence of Novel Behaviours from Complex Non Linear Systems 106
Lorenzo Riano and T.M. McGinnity
GRAVA – Context Programming 113
Paul Robertson and Robert Laddaga
Implementing First-Order Variables in a Graphical Cognitive Architecture 119
Paul Rosenbloom
Cathexis: An Emotional Basis for Human-Like Learning 125
Michael Sellers
Biological and Psycholinguistic Influences on Architectures for Natural
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

Language Processing 131

John F. Sowa
A Bio-Inspired Model for Executive Control 137
Narayan Srinivasa and Suhas E. Chelian
Neural Symbolic Decision Making: A Scalable and Realistic Foundation for
Cognitive Architectures 147
Terrence C. Stewart and Chris Eliasmith
The Role of the Basal Ganglia – Anterior Prefrontal Circuit as a Biological
Instruction Interpreter 153
Andrea Stocco, Christian Lebiere, Randall C. O’Reilly and
John R. Anderson
Learning to Recognize Objects in Images Using Anisotropic Nonparametric
Kernels 163
Douglas Summers-Stay and Yiannis Aloimonos

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 xiii

Disciple Cognitive Agents: Learning, Problem Solving Assistance, and

Tutoring 169
Gheorghe Tecuci, Mihai Boicu, Dorin Marcu and David Schum
Attention Focusing Model for Nexting Based on Learning and Reasoning 170
Akshay Vashist and Shoshana Loeb
A Neurologically Plausible Artificial Neural Network Computational
Architecture of Episodic Memory and Recall 175
Craig M. Vineyard, Michael L. Bernard, Shawn E. Taylor,
Thomas P. Caudell, Patrick Watson, Stephen Verzi, Neal J. Cohen
and Howard Eichenbaum
Validating a High Level Behavioral Representation Language (HERBAL):
A Docking Study for ACT-R 181
Changkun Zhao, Jaehyon Paik, Jonathan H. Morgan and Frank E. Ritter

Manifesto

Introducing the BICA Society 191

Alexei V. Samsonovich, Kamilla R. Jóhannsdóttir, Andrea Stocco and
Antonio Chella

Review

Toward a Unified Catalog of Implemented Cognitive Architectures 195

Alexei V. Samsonovich

Subject Index 245

Author Index 247
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 This page intentionally left blank
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 Conference Papers and Extended Abstracts
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 This page intentionally left blank
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 Biologically Inspired Cognitive Architectures 2010 3
A.V. Samsonovich et al. (Eds.)
IOS Press, 2010
© 2010 The authors and IOS Press. All rights reserved.
doi:10.3233/978-1-60750-661-4-3

Reverse Engineering the Vision System

James ALBUS
Krasnow Institute for Advanced Studies, George Mason University
4400 University Drive MS 2A1, Fairfax, VA 22030-4444, USA
[email protected]

Abstract

The vision system is perhaps the most well understood part of the neocortex. The input
from the eyes consists of a set of images made up of pixels that are densely packed in
the fovea and less so in the periphery. Each pixel is represented by a vector of
attributes such as color, brightness, spatial and temporal derivatives. Pixels from each
eye are registered in the lateral geniculate nucleus and projected to the cortex where
they are processed by a hierarchy of array processors that detect features and patterns
and compute their attributes, state, and relationships. These array processors consist of
Cortical Computational Units (CCUs) made up of cortical hypercolumns and their
underlying thalamic and other subcortical nuclei. Each CCU is capable of performing
complex computational functions and communicating with other CCUs at the same and
higher and lower levels. The entire visual processing hierarchy generates a rich,
colorful, dynamic internal representation that is consciously perceived to be external
reality. It is suggested that it may be possible to reverse engineer the human vision
system in the near future [1].

References
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

[1] J.S. Albus, Reverse Engineering the Brain, International Journal of Machine Consciousness 2 (2010),
193-211.

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 4 Biologically Inspired Cognitive Architectures 2010
A.V. Samsonovich et al. (Eds.)
IOS Press, 2010
© 2010 The authors and IOS Press. All rights reserved.
doi:10.3233/978-1-60750-661-4-4

Application Feedback in Guiding a Deep-

Layered Perception Model
Itamar Arela and Shay Berantb
a
Department of Electrical Engineering & Computer Science, University of Tennessee
b
Binatix, Inc., Palo Alto, CA

Abstract. Deep-layer machine learning architectures continue to emerge as a

promising biologically-inspired framework for achieving scalable perception in
artificial agents. State inference is a consequence of robust perception, allowing
the agent to interpret the environment with which it interacts and map such
interpretation to desirable actions. However, in existing deep learning schemes, the
perception process is guided purely by spatial regularities in the observations, with
no feedback provided from the target application (e.g. classification, control). In
this paper, we propose a simple yet powerful feedback mechanism, based on
adjusting the sample presentation distribution, which guides the perception model
in allocating resources for patterns observed. As a result, a much more focused
state inference can be achieved leading to greater accuracy and overall
performance. The proposed paradigm is demonstrated on a small-scale yet
complex image recognition task, clearly illustrating the advantage of incorporating
feedback in a deep-learning based cognitive architecture.

Keywords. Deep-layered machine learning, perception, spatiotemporal inference.

Introduction

Perception is at the core of intelligent systems. The vast amount of information that
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

humans (and advanced robotic systems) are exposed to every second of the day is
driven by sensory inputs that span a huge observation space. The latter is due to the
natural complexity of the world with which such systems interact. This inestimable
amount of information must be somehow efficiently represented if one is to
successfully function in the real-world. Deep machine learning (DML) is an emerging
field [1] within cognitive computing which may be viewed as a framework for
effectively coping with vast amounts of sensory information.
One of the key challenges facing the field of cognitive computing is perception in
high-dimensional sensory inputs. An application domain in which this challenge clearly
arises is pattern recognition in large images, where an input may comprise of millions
of pixels. These millions of simultaneous input variables span an enormous space of
possible observations. In order to infer the content perceived, a system is required to
map each observation to a possible set of recognized patterns. However, due to a
phenomenon known as the curse of dimensionality [2], the complexity of training a
system to map observations to recognized pattern classes grows exponentially with the
number of input variables. Such growth primarily pertains to the number of examples
the system is required to be presented with prior to becoming adequately proficient.

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 I. Arel and S. Berant / Application Feedback in Guiding a Deep-Layered Perception Model 5

A common approach to overcome the curse of dimensionality is to pre-process the

data in a manner that reduces its dimensionality to such a level that can be effectively
processed by a classification module, such as a multi-layer perceptron (MLP) artificial
neural network. Such dimensionality reduction is often referred to as feature extraction.
Its goal is to retain the key information needed to correctly classify the input within a
lower-dimensional space. As a result, it can be argued that the intelligence behind
many pattern recognition systems has shifted to human-engineered feature extraction
processes, which at times are very difficult and highly application-dependent.
Moreover, if incomplete, distorted or erroneous features are extracted classification
performance may degrade significantly.
Some recent neuroscience [3][4] findings have provided insight into the principles
governing information representation in the mammal brain, leading to new ideas for
designing systems that represent information. One of the key findings has been that the
neocortex, which is associated with many cognitive abilities, does not explicitly pre-
process sensory signals, but rather allows them to propagate through a complex
hierarchy of modules that, over time, learn to represent observations based on the
regularities they exhibit. Such hierarchical representation offers many advantages,
including robustness to diverse range of noise and distortions in the data, as well as the
ability to cope with missing or erroneous inputs.
DML continues to emerge as a promising, biologically-inspired framework for
complex pattern inference. A key assumption in DML is that representation is driven
by regularities in the observations. As one ascends the hierarchical architecture of
DML systems, more abstract notions are formed. Hence, in higher layers of the
hierarchy scope is gained while detail is lost. This appears to be a pragmatic trade off,
as well as a biologically plausible one. In the context of artificial general intelligence
(AGI) [5], one can view perception as being identical to modeling data, in the sense
that partial observations of a large visual field are utilized in inferring the state of the
world with which the agent interacts.
In most existing deep learning schemes [6][7] there is either none or weak
relationship between the (unsupervised) training of the model (DML) engine and the
decision making modules. This forces DML systems to form a representation purely
based on regularities in the observation rather than being driven also by the application
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

at hand (e.g. visual pattern recognition). It is well known, for example, that neurons in
layer IV of the neocortex receive all of the synaptic connections from outside the cortex
(mostly from thalamus), and themselves make short-range, local connections to other
cortical layers. This suggests that learning may not be driven exclusively by
regularities in the observations, but rather co-guided by external signals.
In this paper we present an elegant methodology for guiding the representation of a
DML system such that it serves as a more relevant perception engine, yielding
improved classification accuracy. The approach is based on adjusting the DML sample
presentation distribution as it is trained such that relevant salient features can be
hierarchically captured.
The rest of this paper is structured as follows. In section 1 we outline the proposed
deep learning system and its operational modes. Section 2 describes the proposed
feedback-based scheme for guiding DML representation. Section 3 describes the
simulation results while in Section 4 conclusions are drawn.

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 6 I. Arel and S. Berant / Application Feedback in Guiding a Deep-Layered Perception Model

1. Deep-layered Inference Engine

The proposed DML architecture comprises of a hierarchy of multiple layers each

hosting a set of identical cortical circuits (or nodes), which are homogeneous to the
entire system, as illustrated in Figure 1. Each node models the inputs it receives in an
identical manner to all other nodes. This modeling, which can be viewed as a form of
lossy compression, essentially represents the inputs in a compact form that captures
only the dominant regularities in the observations. The system is trained in an
unsupervised manner by exposing the hierarchy to a large set of observations and
letting the salient attributes of the inputs be formed across the layers. Next, signals are
extracted from this deep-layered inference engine to a supervised classifier for the
purpose of robust pattern recognition. Robustness here refers to the ability to exhibit
classification invariance to a diverse range of transformations and distortions, including
noise, scale, rotation, displacement, etc.

Deep-layer
Inference Network
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

Figure 1. Deep-layered visual perception network comprising multiple layers hosting identical cortical
circuits. The lowest layer of the hierarchy receives raw sensory inputs. Features generated by the cortical
circuits are passed as input to a supervised classifier.

The internal signals of the cortical circuits comprising the hierarchy may be
viewed as forming a feature space, thus capturing salient characteristics of the
observations. The top layers of the hierarchy capture broader, more abstract, features of
the input data, which are often most relevant for the purpose of pattern recognition.
The nature of this deeply-layered inference architecture involves decomposing
high-dimensional inputs into smaller patches, representing these patched in a compact
manner and hierarchically learning the relationships between these representations
across multiple scales. The underlying assumption is that input signal proximity is
coherent with the nature of the data structure that is being represented. As an example,
two pixels in an image, which are in close proximity, are assumed to exhibit stronger
correlation than that exhibited by two pixels that are very distant. This assumption

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 I. Arel and S. Berant / Application Feedback in Guiding a Deep-Layered Perception Model 7

holds firmly for many natural modalities, including natural images and videos, radar
images and frequency components of an audio segment. In a face recognition
application, for example, the output of the classifier may be a single value denoting
whether or not the input pattern corresponds to a particular person.
DML combined with a classifier may be viewed as a general semi-supervised
learning framework. Training the system can be generalized as follows. During the first
step, a set of unlabeled samples (i.e. inputs/observations that do not have a known class
label associated with them) are provided as input to the DML engine. The latter will
learn from such samples about the general structure of the sensory input space it is
presented with. During the second step, a set of labeled samples (i.e. inputs that have a
distinct class labels associated with them) is provided as inputs. Signals are extracted to
a classifier, which is then trained in a supervised manner on the labeled set. Testing is
then achieved by presented unseen observations and evaluating the output of the
classifier relative to the actual image class.

2. Application Feedback for Improved Perception

As described above, the semi-supervised framework that applies to most deep machine
learning schemes implies a strict decoupling between the model (i.e. unsupervised
training of the DML architecture) and the application (i.e. classifier). Learning to
represent the input space based purely on regularities in the observations appears
elegant. However, a more pragmatic approach is to guide the learning process of the
DML engine so that it is of greater relevance to the classifier. For example, if the
observations exhibit regularities which are not pertinent to the classifier to perform
well, they may as well be ignored or discarded by the model. Thus, we propose a
feedback mechanism between the classifier and the deep learning engine such that
representation is optimized for the classification process.
The feedback mechanism proposed involves adapting the sample (i.e. observation)
presentation distribution to the DML engine based on results obtained from the
classification process. To do so, the DML and classifier trainings are performed
concurrently, rather than in succession. This is somewhat of a paradigm shift from
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

existing DML methodologies, but one that is argued vital. The classifier considered is a
simple MLP feedforward neural network. As opposed to uniformly presenting samples
from the unlabeled set to the DML engine, the sample presentation distribution is
modified such that observations which need to be reinforced are presented more
frequently. The need to reinforce presentations is derived from the classification error
measured such that observations (i.e. input samples) that yield relatively high errors
will be more frequently presented to the DML engine.

3. Simulation Results

The simulation results pertain to a simple image classification scenario. The goal is to
provide an example highlighting the advantage of adjusting the sample distribution in
an online manner. A database consisting of a train and test sets of images was created
synthetically. Each of the two set contained 500 images, belonging to 9 classes (the
letters 'C','G','H','M','P','R','T','X','Z'). These classes were each represented by a template
image. Every image in both sets of the database was created from one of the template

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 8 I. Arel and S. Berant / Application Feedback in Guiding a Deep-Layered Perception Model

images, with random distortion applied. The distortion included scaling, rotation,
erosion and application of additive noise. Sample images from the test set are
illustrated in Figure 2.

Figure 2. Examples of letter images distorted randomly and used in evaluating the proposed DML system.

The system was trained in a supervised manner on the training set, whereby the
classifier was being targeted with a vector filled with '0's except for a single '1' in the
location that corresponds to the image label. Testing was conducted on test set images,
which are guaranteed to differ from the training set.
During the training phase, a model and a MLP neural network classifier were
trained concurrently, in two modes. In the first mode, there was no classifier-model
feedback involved, and the sample distribution remained uniform throughout the entire
learning process. In the second mode, a feedback mechanism was applied in order to
influence the sample presentation distribution, such that images with high classification
error were presented more frequently, as described above.
In both modes, training was performed in batches. During each batch, 100 images
were randomly preselected for presentation from the N images comprising the training
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

set. In the first mode, these images were selected uniformly and independently at
random. During the second mode, an adaptive presentation scheme was applied. At the
end of each batch, after the DML parameters update, the classification error for each of
the images in the training set has been evaluated. As a result, the sample presentation
distribution was updated by applying a simple convex summation of the form

݁‫݅ݐ‬
‫ݐ݌‬൅ͳ ൌ ߙ‫ ݅ݐ݌‬൅ ሺͳ െ ߙሻ ǡ ݅ ൌ ͳǡ Ǥ Ǥ ǡ ܰ‫ ݐ‬ൌ ʹǡ͵ǡ ǥ (1)
݅ σ݅ൌͳ ݁‫݅ݐ‬
ܰ

where ‫݌‬௜௧ =1/N denotes the probability of selecting image i for presentation at batch t, ݁௜௧
the classification error for the ith image (calculated as the element mean on the absolute
difference between the classifier output and the target vector) at batch t, and 0<α<1 is
an coefficient set to 0.95.
Figure 3 illustrates the classification accuracy (on both the training and testing
sets) as a function of the number of batches. Classification did not reach 100%
accuracy intentionally, as the number of training samples was limited and the
distortions vast. This was chosen in an aim to accentuate the difference of applying the

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 I. Arel and S. Berant / Application Feedback in Guiding a Deep-Layered Perception Model 9

feedback mechanism on a simple task. As can be observed, on both training and testing
sets, performance was consistently higher when the proposed feedback mechanism was
applied, suggesting that guiding the representation of the DML engine by emphasizing
underperforming samples leads to improved overall classification performance.
1

0.9
Normalized accuracy

0.8

0.7

Train set without feedback

0.6 Test set without feedback
Train set with feedback
Test set with feedback
0.5

0.4
200 300 400 500 600 700 800 900 1000
Batches performed

Figure 3. Classification accuracy as a function of the number of batches for both train and test set images.

4. Conclusions

In stark contrast to mainstream schemes, this paper presents a natural way in which
data presentation in deep machine learning systems can be driven by the application,
rather than purely by regularities in the observations. An online technique for adjusting
the sample distribution based on classification error retains the hands-off attributes of
DML, as it requires no human intervention. Simulation results clearly illustrate the
benefits of a feedback mechanism from the application to the DML model. Moreover,
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

the proposed approach can be broadly applied to other DML architectures and
application domains.

References

[1] Y. Bengio, Learning Deep Architectures for AI, Foundations and Trends in Machine Learning, 2,(2009),
1-127.
[2] R. Bellman, Dynamic Programming, Princeton University Press, 1957.
[3] T. Lee, D. Mumford, R. Romero, V. Lamme, The Role of the Primary Visual Cortex in Higher Level
Vision, Vision research, 38 (1998), 2429-2454.
[4] G. Wallis, H. Bülthoff, Learning to recognize objects, Trends in Cognitive Sciences, 3 (1999), 23-31.
[5] I. Arel, S. Livingston, Beyond the Turing Test, IEEE Computer, 42 (2009), 104-105.
[6] M. Ranzato, F.J. Huang, Y. Boureau, Y. LeCun, Unsupervised Learning of Invariant Feature Hierarchies
with Applications to Object Recognition, Proc. Computer Vision and Pattern Recognition Conference,
2007.
[7] G. E. Hinton, S. Osindero, Y. Teh. A Fast Learning Algorithm for Deep Belief Nets, Neural Computation,
18, (2006), 1527-1554.

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 10 Biologically Inspired Cognitive Architectures 2010
A.V. Samsonovich et al. (Eds.)
IOS Press, 2010
© 2010 The authors and IOS Press. All rights reserved.
doi:10.3233/978-1-60750-661-4-10

NeuroNavigator: A Biologically Inspired

Universal Cognitive Microcircuit
Giorgio A. ASCOLIa,b,1 and Alexei V. SAMSONOVICH b
a
Center for Neural Informatics, Structures, and Plasticity
b
Molecular Neuroscience Department, Krasnow Institute for Advanced Study
George Mason University, Fairfax, Virginia 22030-4444,USA
[email protected], [email protected]

Abstract. We present a universal building block for cognitive machines, called

NeuroNavigator, inspired by theories of the hippocampus. The module is designed
to fit both biological plausibility and constraints of forthcoming neuromorphic
hardware. Its functions may range from spatial navigation to episodic memory
retrieval. The goal of the present study of NeuroNavigator is to show the
scalability of the model. The core of the architecture is based on our previously
described model of hippocampal function and includes 3 layers (DG, CA3, CA1)
of spiking neurons with noisy STDP synaptic connections among neighboring
layers. The model is applied to a spatial navigation paradigm in a hierarchical
virtual environment, the metrics of which need to be learned by exploration. The
goal in each trial is set arbitrarily as any one of the previously seen objects or
features. In order to navigate toward the goal, the agent needs to “imagine”
previously performed available moves at the current location and select one of
them, using the acquired spatial knowledge. This process controlled by
NeuroNavigator is repeated until the goal is reached. Overall, the simulation
results show robustness and scalability of the solution based on a biologically-
inspired network of spiking neurons and STDP synapses.

Keywords. Hippocampus; spatial learning; spiking networks

Copyright © 2010. IOS Press, Incorporated. All rights reserved.

Introduction

Recently developed and rapidly advancing neuromorphic hardware technology [4]

allows for implementation of large networks of spiking neurons and spike-time-
dependent-plasticity (STDP) synapses, opening broad possibilities for implementation
of large-scale neuronal systems with mammal-level intelligent capabilities. As a first
step, it is critical to demonstrate that a simple neural network model compatible with
the new hardware (a) can perform a useful cognitive function that supports intelligent
behavior in a real or virtual environment, and (b) is scalable from dozens to a million of
neurons with virtually all-to-all (up to 1010) synaptic connections, up to 10,000
connections per neuron (the network size afforded by the new technology).
Many cognitive and learning functions require a solution of the problem of
navigation in a previously explored environment. Because of the generality of this task,
a specialized hardware microcircuit capable of solving it in a scalable manner will

1
Corresponding Author.

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 G.A. Ascoli and A.V. Samsonovich / NeuroNavigator 11

become a universal building block for cognitive artifacts capable of autonomous

learning. Many examples of cognitive and learning tasks involve the problem of
navigation in a previously explored discrete space. One is pathfinding, or spatial
navigation, performed in a familiar stationary environment that can be discretized in its
internal representation in the system. Another, less trivial example is planning under
the assumption of complete knowledge of the world, when a plan is represented as a
trajectory in the abstract space of states, most of which were previously explored (e.g.,
planning of routine actions like shopping, etc. on the way home). A yet less trivial
example is strategic episodic memory retrieval, when the target episode is given,
however, in order to understand it, one needs to retrieve memories of all related key
episodes [1]. Further examples include bootstrapped learning, logical problem solving,
and more. We have previously demonstrated [1] that one and the same, biologically-
inspired model is capable of solving all tasks of this broad subcategory of search
problems. The solution applied to large-scale tasks becomes most efficient and
practically attractive, when implemented in a specialized neuromorphic hardware based
on spiking neurons and STDP synapses. The cognitive architecture of NeuroNavigator
is constrained by the hardware parameters and by minimal requirements of biological
plausibility. Therefore, the outcome is expected to be both practically useful and
realistic as a model of brain functions.

1. Materials and Methods

1.1. Description of the Paradigm

Our task in this study is to design, implement and study in simulations a model system
of up to 1,000,000 neurons capable of solving spatial and cognitive tasks. In particular,
we demonstrate in simulations a combination of features and phenomena including
network activity, network stability, synaptic plasticity, self-organization in response to
sensory stimulation and modulation/reinforcement, unsupervised learning, and
reinforcement learning. Specifically, the paradigm includes the following elements.
The agent is embedded in a hierarchical virtual environment, elements of which
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

are indoor environments, cities, and a highway network (Figure 1). Indoor
environments may be implemented as a graph of available locations and transitions
between them (e.g., Figure 1 C) or continuous 3D environments with a virtual robot
body embedded in them (implemented using CASTLE simulator [8]: Figure 1 D).
Higher-level environments are implemented in discrete versions only. Each discrete
environment may include up to N locations, one of which (location number 1) is the
entry to the environment. Continuous indoor environments may include up to K
landmarks that the robot can reach.
The behavioral paradigm includes two phases: exploration and navigation. During
exploration, the robot moves randomly in the hierarchy of discrete environments,
exploring all locations and most available transitions in discrete environments, and
associating visible objects and features with locations. The robot avoids as much as
possible making U-turns. In continuous environments, the robot follows either the sight
of a randomly chosen visible landmark or the guide (an avatar used as a visual cue to
guide the robot) to perform exploratory motion among landmarks. During this process,
the system learns the set of landmarks and available transitions between them, as well
as features associated with each landmark.

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 12 G.A. Ascoli and A.V. Samsonovich / NeuroNavigator

During navigation, the agent is placed at a random, previously visited location and
is given a goal in the form of a previously seen feature or object that needs to be
reached in the environment. NeuroNavigator plans steps toward the goal, and the
virtual robot performs the steps. In a discrete environment, a step is a move along a link
of the graph. In a continuous environment, a step involves navigation toward a visible
landmark using the visual cue and motor control (implemented outside of
NeuroNavigator). The agent performance is evaluated by the percent of correct
solutions and by the average time / path length used for navigation compared to the
shortest path, as functions of the time spent during exploration of the environment. The
challenge consists in designing and implementing a network of spiking neurons capable
of performing the described paradigm.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

Figure 1. Example views of the hierarchy of environments. A: highway network, B: city, C: discrete indoor
environment, D: continuous 3D indoor environment with a virtual robot in it, E: the entire hierarchy of
environments.

1.2. Design of NeuroNavigator

In this work we argue that an efficient, scalable solution of the challenge outlined
above can be found in the form of a biologically plausible network of spiking neurons
of Izhikevich type [2] or simple integrate-and-fire neurons, connected by synapses that
obey STDP (spike-time-dependent plasticity) learning rules [3].
We start by consideration of a problem of spatial navigation in a small discrete
environment, and show how a network of spiking neurons learns to navigate in this
space. Then we move to a scalable solution of the challenge involving the large
hierarchical environment, with discrete and continuous elements.

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 G.A. Ascoli and A.V. Samsonovich / NeuroNavigator 13

1.2.1. Architecture Components

The architecture of a NeuroNavigator module is represented in Figure 2. The core
includes neuronal layers labeled DG (dentate gyrus), CA3, CA1, and Perirhinal, among
which CA3 and CA1 layers are divided into the 4 submodules associated with the 4
abstract directions of available moves labeled West, East, North and South2. In addition
to the core, several interface layers of neurons are used that could be implemented as
real neurons in the hardware or virtual neurons in the simulator connected to the
neuromorphic hardware. These interface neurons include layers labeled Exploration,
Navigation (both representing the current location of the robot), Move Direction
(works during exploration only), Feature/Goal, and Output that tells the direction of the
selected move at each step during navigation.

1.2.2. Architecture Dynamics

In this subsection we explain architecture dynamics at the top level. During exploration,
neurons representing the current location are excited in all five shown populations by
processed sensory input (red arrows: Figure 2 A). As the agent moves continuously
from one location to another, DG-to-CA3 connections are formed by associative STDP
between neurons representing neighboring locations. This happens, however, only in
those submodules that correspond to the current direction of motion, due to the
modulation of synaptic plasticity. At the same time, long-range connections between
CA3 neurons representing recently visited locations and CA1 neurons representing the
currently active goal location are potentiated by STDP in all submodules.
During navigation, the agent mentally explores available moves and selects one
that causes the strongest response in the goal cell (Figure 2 B).

1.3. Implementation on a Supercomputer

Before the model can be implemented in the neuromorphic hardware, we need to test it
on a parallel supercomputer (the Server), which implies, in particular, the usage of a
special language PyNN for defining the model. This language allows for specification
of a large number of architectural elements at once, with probabilistic models
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

describing detailed connectivity of the network. Therefore, in order to create one-to-

one laminar connections from the model input-output units representing spatial
locations to the corresponding DG, CA3 and CA1 units (figure 2), a self-organization
process can be used based on winner-take-all dynamics and associative STDP.
Similarly, one-to-one connections can be established between the model perirhinal
units and the input-output feature units.

2. Results: Addressing Specific Challenges

The main aspect of the present challenge is not to find an efficient hardware solution
for a cognitive problem, but to demonstrate that a range of interesting cognitive
problems can be solved with the new hardware. The nature of the hardware (and also

2
Only East and West submodules are shown in Figure 2. The number 4 is set arbitrarily. Each available
move at each location is assigned a unique abstract direction that may or may not coincide with the
allocentric direction.

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 14 G.A. Ascoli and A.V. Samsonovich / NeuroNavigator

the nature of biological systems that inspired the hardware) suggests that almost all
communications with and within our network must be mediated by spikes. To address
this challenge, we introduce additional layers serving the interface function. In general,
our simulations demonstrate that the design works in selected paradigms. These results
will be presented elsewhere. Here we present our study of specific challenges –
elements of the complete solution. Our key result is the design of the architecture that
allows for all information processing to be performed by propagating spikes (Figure 2).

A Module Perirhinal

West East
CA1 Legend:
Neuron
Spiking neuron
Excitiation
Inhibition
Learned
CA3 connection
Plasticity on
Plasticity off

DG

Exploration: Navigation: current Head Feature/Goal

current location location direction
Input: mode, location ID, move direction, perceived feature

B Module C Module: 9 sub-modules

CA1- CA3- CA1-

Perirhinal Parahippocampal West North North

CA3- CA3-
DG
CA1 West East

CA1- CA3- CA1-

South South East
CA3
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

DG

D The head-direction
system: 4 output
Feature/Goal
Exploration: Navigation: Head wires, one active
current location current location direction

Input

Figure 2. Cognitive architecture of NeuroNavigator during A: exploration, B: navigation. Small filled

circles represent neurons. East and West correspond to allocentric directions assigned to sub-modules. C, D:
possible arrangement of the architecture components on a chip in a hardware implementation.

2.1. Effects of synaptic noise

One of the challenges in this study is to build a robust system that can operate reliably
in various environments despite the noisy nature of its elements. In particular, the value
of synaptic efficacy is subject to static and dynamic noise; therefore, meaningful small
differences in the expected synaptic efficacies may not be reliably detectable. An
approach to overcoming this difficulty is to use voting of identical systems working
synchronously in parallel on the same navigation task.

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 G.A. Ascoli and A.V. Samsonovich / NeuroNavigator 15

We checked how the level of static synaptic noise affects the probability of goal
reaching in the virtual city environment (one fragment of the hierarchy). The results are
represented in Figure 3 A. Since the expected level of noise in the hardware is 0.05, we
need voting of parallel modules to solve navigation tasks reliably.

2.2. The additive nature of STDP

The learning rule dictated by the hardware (and also suggested by biology) may
not be optimal for implementation of the selected model. The approach here is, again,
to increase the number of parallel voting modules in order to avoid mistakes in
navigation.
In our simulations (Figure 3 B), the exploration epoch for each module was limited
to 300 steps. The retrieval path was limited to 20 steps. The noised values of synaptic
weights were truncated to the interval (0,1). The voting was done by the majority of
choices for the next step. The expected slope (scaling law) of the curve in panel B is
that of a square root.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

Figure 3. A: The success rate as a function of Gaussian synaptic noise. B: The success rate as a function of
the effect of the number of voting modules. C. Results of exploration of a large hierarchy of environments
after 300,000 exploratory moves: data for the “highway network”. D: Dealing with hubs by augmenting the
design of the network. Module: 24×2+1=49 sub-modules.

2.3. Large exploration space and long shortest paths

Another challenge relates to the scalability of the task and the architecture. As the
space becomes bigger, it should take exponentially longer time to explore all possible

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 16 G.A. Ascoli and A.V. Samsonovich / NeuroNavigator

trajectories in it. Our approach here is to break the large space into a hierarchy of
small-size environments, each of which can be explored and navigated separately. Then
the solution of a navigation problem can be found as a multi-leg path. Again, the
simulation results demonstrate the feasibility of this approach (Figure 3 C).

2.4. Multiplicity of available moves

Because all available moves in our architecture are explored in parallel by separate
submodules, the approach to dealing with this challenge is to increase the number of
submodules assigned to different directions of motion (Figure 3 D). Hubs (i.e., nodes
with large numbers of connections) are typical for brain networks [5] and many real-
world networks in general [6,7].

3. Conclusions

Our task in this study is to design, implement and study in simulations a model system
of up to 1,000,000 neurons capable of solving spatial and cognitive tasks. In particular,
we demonstrate in simulations a combination of features and phenomena including
network activity, network stability, synaptic plasticity, self-organization in response to
sensory stimulation and modulation/reinforcement, unsupervised learning, and
reinforcement learning. The network is constructed from spiking units and shows
scalability in size, level of noise and in topological complexity. Internal network
dynamics resemble those of the brain performing similar cognitive tasks [9].

3.1. Acknowledgments

We are grateful to Drs. Art Pope and Aleksey Nogin for help with technical issues.
This work is supported by DARPA DSO Grant SyNAPSE, subcontract number
801887-B8.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

References

[1] A. V. Samsonovich and G. A. Ascoli, A simple neural network model of the hippocampus suggesting its
pathfinding role in episodic memory retrieval, Learning & Memory 12 (2005): 193–208.
[2] E.M. Izhikevich and G.M. Edelman, Large-scale model of mammalian thalamocortical systems, Proc. of
the National Academy of Sciences of the United States of America 105 (2008): 3593-3598.
[3] E.M. Izhikewich, Solving the distal reward problem through linkage of STDP and dopamine signaling.
Cerebral Cortex 17 (2007): 2443-2452.
[4] C. Diorio, P. Hasler, B.A. Minch, and C. Mead, A complementary pair of four-terminal silicon synapses,
Analog Integrated Circuits And Signal Processing 13 (1997): 153-166.
[5] C. J. Honey, R. Kötter, M. Breakspear and O. Sporns, Network structure of cerebral cortex shapes
functional connectivity on multiple time scales, Proceedings of the National Academy of Sciences of the
United States of America, 104(2007): 10240-10245.
[6] A.L. Barabasi and R. Albert, Science 286 (1999): 509.
[7] A. L. Barabasi and E. Bonabeau, Scale-free networks, Scientific American 288 (2003): 60–69.
[8] A. Pope and P. Langley, CASTLE: A framework for integrating cognitive models into virtual
environments, in A. V. Samsonovich, Biologically Inspired Cognitive Architectures: Papers from the
AAAI Fall Symposium, AAAI Technical Report FS-08-04 (152), AAAI Press, Menlo Park, CA, 2008.
[9] J.J. Hopfield, Neurodynamics of mental exploration, Proceedings of the National Academy of Sciences
of the United States of America, 107(2010): 1648-1653.

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 Biologically Inspired Cognitive Architectures 2010 17
A.V. Samsonovich et al. (Eds.)
IOS Press, 2010
© 2010 The authors and IOS Press. All rights reserved.
doi:10.3233/978-1-60750-661-4-17

BINAReE:
(Bayesian Integrated Neural Architecture for Reasoning and Explanation)

Robert (Rusty) Bobrowa, Paul Robertsonb and Robert Laddagab

a
Raytheon BBN Technologies, 10 Moulton Street, Cambridge, MA 02138, USA
b
Dynamic Object Language Labs, Haverhill, MA 01832, USA
a
[email protected], [email protected], [email protected]

Abstract. Our goal is to lay a foundation for a model of high level cognition in
humans that respects biological constraints. To do so, we integrate realistic neural
models of attention, memory and control into an active information foraging
system comprising many of the component neural systems and strategies in
published models of high-level active vision. The direction of attentional focus
requires integration of bottom-up and top-down attention. We base our modeling
methodology for both individual components and the overall system operation on
structured hierarchical Bayesian models. These models are built upon
computationally tractable, neurally implementable sampling approximations to
inference and control.

Keywords. Neuroscience, active vision, MCMC, DBN, POMDP, cognitive model

1. Introduction

BBN has proposed developing a system: BINAReE: Bayesian Integrated Neural

Architecture for Reasoning and Explanation. The goal of BINAReE is to obtain real
insight into the cognitive processes involved in sensemaking and at the same time to
build an ambitious and unprecedented integration of computational models from
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

neuroscience. BINAReE integrates realistic neural models of attention, memory and

control into an active information foraging system comprising many of the component
neural systems and strategies in published models of high-level active vision. While
vision by itself is not enough to provide high level cognition, vision is among the most
heavily studied parts of brain function. High-level vision involves at least the Parietal,
Temporal, and Pre-frontal Cortices, Basal Ganglia, and Anterior Cingulate Cortex
(ACC). Active sensemaking based on notions of active vision is the starting point and
a significant innovation of our approach.
BINAReE implements an active information seeking process in which attention is
directed to acquire and integrate new information. The direction of attentional focus
requires integration of bottom-up and top-down attention. We integrate detailed
computational models and experimental evidence localizing both types of control.
We base our modeling methodology for both individual components and the
overall system operation on structured hierarchical Bayesian models [Tenenbaum,
1999]. These models are built upon computationally tractable, neurally implementable
sampling approximations to inference and control. This novel approach mathematically
characterizes the structure of the context frames which are the basis for sensemaking,
and this allows us to model frame learning with established computational techniques

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 18 R. Bobrow et al. / Bayesian Integrated Neural Architecture for Reasoning and Explanation

for hierarchical Bayesian learning. Additionally, sampling algorithms applied to

hierarchical, structured Bayesian models of sensemaking account for a number of
cognitive biases, [Vul, et al, 2009] and can be mapped to neural computations.
The BINAReE project is bound by some common themes: A view that much of the
computational nature of brain function is best modeled in terms of Bayesian inference
processes; a recognition that many functions, especially in sense making, can be
usefully thought of in terms of a chain of decisions and that Partially Observable
Markov Decision Processes (POMDP) provides a powerful approach to frame the
endeavor; and finally, that the neuroscience, the modeling, and the implementation can
be brought together around the notion of Monte Carlo sampling methods. There is
much support for sampling-based approaches in the neuroscience literature [Goodman
et al., 2008; Griffiths & Tenenbaum, 2006; Vul et al., 2009, Fiser et al. 2010] and
today’s hardware, especially GPUs, offer the ability to build parallel implementations
whereby non-trivial high-level (not neuron-level) modeling experiments can be
constructed and tested to yield a driver for neuroscience.

2. Overview of the Approach

In order to summarize the nature and benefits of our approach, we begin here by
walking through, at a high level, the early stages of our proposed research agenda. The
walk through begins with selection of a frame, then shows how information foraging in
the context of that frame is abstractly implemented by our component models of seven 1
brain regions. We begin by exploring the instantiation of partially grounded frames
(PGFs) and conclude the summary by continuing the story in a little less detail in order
to paint an overview of the sensemaking loop and its relationship to the brain areas of
interest to this program and the models and team members that form our team.
2.1. Harbor Island Example Sensemaking Process
Consider how we would apply sensemaking to GEOINT data from Harbor Island,
Seattle, as in Figure 1. The sensemaking task at hand is to make a set of determinations
about the scene, constructing a hierarchical description using a set of nested frames
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

based on (and grounded in) the entities observed in the Urban Feature Data layer.
These frames describe a generative model for the observations: (1) At the highest level
is the description of the
scene as a port terminal:
this partially grounded
frame (PGF) includes (2) a
number of subsidiary
frames for port
components and their
spatial relations. (3) Each
of these subsidiary frames
is directly grounded in the
immediately observed
entities, such as a number Figure 1. Harbor Island layers. Raw IMINT at the bottom, extracted
of storage tank farms, a feature layer in the middle, and the hierarchical frame constructed
by the sensemaking model at the top.

1
IT, PFC, ACC, BG, SC, P-LIP, BNMS

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 R. Bobrow et al. / Bayesian Integrated Neural Architecture for Reasoning and Explanation 19

group of linearly arranged ship berths along the shoreline, a container holding area, etc.
The goal of sensemaking is to infer the appropriate frames and their properties at each
level of the hierarchy, thus constructing a complete, description of the scene as a whole.
We begin with a root piece of data, such as a geospatial layer and some sense of
what we are looking for. From this root, the sensemaking process will cycle between
information foraging based on previously gathered leads, fitting the new data to the
evolving story, and deciding on further actions.
The first step is to instantiate a driver for further sensemaking by analyzing the
provided data to estimate the global frame. The gist [Oliva, Torralba, 2006] of a scene
computed from the coarse low-frequency statistics of an image can be used to estimate
the scene type. The basic model suggests that the primary areas of the computation
occur in the infero-temporal cortex (IT) and that the resulting gist of the scene forms a
global structure which supports the process of data collection and sensemaking in the
prefrontal cortex (PFC).
A global frame is selected from the set of previously learned frames by weighting
each frame by its prior probability as well as the match of the “gist” features computed
from low-frequency information from the scene. Thus this process instantiates a
partially grounded frame (PGF) which provides unbound slots, or terminals – initially
these are unknown and unfilled, but this frame represents the current state of
knowledge, and supports information about subsequent decision making. Our
implementation of Torralba’s models uses a software capability called GRAVA
[Robertson, Laddaga, 2003] that jointly represents functional models, their
communication pathways, and local memory.
Our example so far brings in two areas, IT and PFC, a form of Bayesian inference
based on statistical measures of scene content, and a decision process – what to do next
given the ungrounded parts of the PGF, the availability, and cost of obtaining
additional information. Figure 2 shows a simplified view of our architecture.
Data foraging is now controlled by a combination of the anterior cingulate cortex
(ACC) and basal ganglia (BG), based on models developed by Joshua Brown
[Alexander & Brown, 2010]. Elaborating the PGF at this point is a decision process.
The decision to be made is: given the set of ungrounded components of the frame, what
data should be sought next? The ACC and BG select this data based on a model of
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

what is most likely to be successfully grounded next, and its likely impact on the
estimate of the posterior probability of the frame. The result of the decision procedure
would be either an overt
shift of focus attention
(resulting in the superior
colliculus (SC) initiating a
saccade to a different place)
or a covert focus attentional
shift resulting in a refocus on
specific kinds of data.
The flood of data
resulting from the focus shift
now needs to be filtered by
Figure 2. Simplified Architecture Diagram
the attentional system. The
attentional mechanism draws

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 20 R. Bobrow et al. / Bayesian Integrated Neural Architecture for Reasoning and Explanation

upon a recent model of attention by Poggio [Chikkerur, et al, 2009] which allows much
of the data coming into the parietal-LIP, as a result of refocusing, to be ignored. This
attention mechanism is another of the Bayesian processes modeled as a Bayes Network.
The features that survive are compared by the ACC against what was predicted.
The level of progress is communicated to the brainstem neuromodulatory system
(BSNMS). If inadequate progress appears to be being made the BSNMS forces an
adjustment of focus to explore a different area of interest in order to further ground the
frame.
This loop of information foraging continues until enough of the frame has been
grounded. Meanwhile, the learning of context frames is implemented through
interaction of the medial temporal lobe (MTL) with the basal ganglia and prefrontal
cortex. This process is modeled as the learning of a hierarchical Bayesian model.
The sensemaking process described above draws upon neuroscience models and
draws heavily upon Bayesian inferencing, Markov decision processes, and the
grounding of frames in working memory while new frames are learned and deposited
into long term memory.

3. Related Research

A number of neural architectures have been proposed for describing integrated function
of the brain by designing learning rules and dynamics from bottom-up constraints on
the function of individual neurons and their interconnections.
Leabra or "Local, Error-driven and Associative, Biologically Realistic Algorithm"
is a model of learning which combines Hebbian and error-driven learning. It is used to
mathematically predict outcomes based on inputs and previous learning influences.
[O'Reilly, 1996].
Adaptive Resonance Theory (ART) is a theory of brain information processing
developed by Stephen Grossberg and Gail Carpenter [Grossberg & Carpenter, 2003]. It
includes neural network models using supervised and unsupervised learning methods,
and addresses pattern recognition and prediction, among other issues. In ART, object
identification and recognition involves the interaction of “top-down” observer
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

expectations with “bottom-up” sensory information.

Temporal Difference (TD) Learning is an approach to learning to predict
quantities based on future values of a signal [Tesauro, 1995]. It is mostly used for
reinforcement learning. TD learning combines Monte Carlo (learns by sampling) and
dynamic programming (bootstrapping) methods. It is a supervised learning process in
which the training signal for a prediction is a future prediction.
Although such bottom-up integrated neural modeling approaches inform the
mechanisms that may carry out various computations in the brain; they do not directly
constrain high-level computations like inference about geospatial structures and
information-gathering decisions. Our goal is to complement these approaches by
starting with top-down computational and algorithmic constraints and to connect these
processes to neural learning models. We aim to build an integrated model that
describes sampling over joint, hierarchical models of the world and the task. This
approach connects high-level tasks to information foraging and meta-cognitive
decisions about which evidence to seek out, which data to remember, and how to plan
future actions.

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 R. Bobrow et al. / Bayesian Integrated Neural Architecture for Reasoning and Explanation 21

4. The System Architecture

BINAReE will leverage GRAVA as its integrative infrastructure. GRAVA was

originally designed as an architecture for mobilizing contextual information for solving
interpretation problems [Robertson & Laddaga, 2003] especially in image analysis
[Robertson & Brady, 1999]. The structure of GRAVA is ideal as an integrative
framework for a number of reasons: (1) The approach taken by GRAVA of identifying
contexts and then invoking agents that are suited to operating in those contexts maps
easily onto the BINAReE model of identifying PGFs and then foraging for data to add
additional grounding; (2) GRAVA permits implementations of models to be provided
and for explicit communication and control connections to be defined; (3) The
architecture allows for models to be implemented in a hierarchical fashion allowing for
both peer and parent modules to be defined and connected; (4) Support for working
memories local to each model are provided for in the infrastructure; (5) GRAVA
supports Bayesian decision processes.
The principles of GRAVA are that most interpretation tasks are not simple
homogeneous computational problems, but rather involve the careful integration of
information from a wide range of sources with different representations and with
different costs associated with their use. The costs may be computational costs or
attentional costs – more generally, resource allocation problems. The problem of
sensemaking in the brain fits exactly that description and it is helpful, when thinking
about natural sensemaking computationally, to consider these issues explicitly.

5. Evaluation: Accounting for Human Biases in Reasoning

In our integrated model, cognitive biases arise due to the interaction of specific
hierarchical Bayesian models within which: 1) reasoners are performing a task, 2)
models of the task determine information seeking policies and strategies, and 3)
sample-based approximate inference algorithms approximate these computations.
Anchoring and adjustment refers to the fact that human estimation is drawn to an
initial lure, and then slowly adjusts away from this hypothesis. We believe this bias
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

arises from the sampling algorithms that the brain employs in specific areas: Markov
chain Monte Carlo (MCMC). MCMC refers to a family of algorithms that sample from
the posterior distribution of complex statistical models by gradually drifting through
the hypothesis space of interpretations– a process that describes the dynamics of human
bistable perception [Gershman, Vul, Tenenbaum, 2010] and the stochastic neural
dynamics in cases of bistable inputs [Moreno-Bote, et al, 2009]. When people exhibit
anchoring and adjustment behavior, they are initially presented with a lure that they
appear to entertain as a plausible hypothesis for the estimate they are trying to achieve.
If this estimation is carried out by MCMC, then the initial anchoring by the lure serves
as the initialization of the Markov chain. After that, the chain will follow a random
walk that appears to “adjust” from the anchor by being biased towards higher
probability regions of the hypothesis space. Thus, anchoring and adjustment will arise
naturally in situations where people are making sense of data by approximating
complex statistical models using MCMC-like algorithms that, due to time or resource
constraints, do not drift sufficiently far from the initialization point.
Similar explanations can be given to address: confirmation bias in information
gathering; change blindness; persistence of discredited evidence; and other biases.

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 22 R. Bobrow et al. / Bayesian Integrated Neural Architecture for Reasoning and Explanation

6. Conclusion

BINAReE is a somewhat realistic high level model of human cognition, based largely
on insights about the operation of the human visual system, and on the structured
Bayesian reasoning approach. It is constrained by modern neuroscience results, and
implemented via an agent based architecture. It heavily uses the separation of context
from attention directed reasoning. BINAReE is able to simultaneously solve significant
real-world problems, and yet represent human-like reasoning biases, as a result of
specific computational processes.

References

Alexander WH, and Brown JW [2010]. Competition between learned reward and error outcome
predictions in anterior cingulate cortex. Neuroimage 49:3210-3218.
Carptenter, G.A and Grossberg, S. [1990] ART: Self-organizing neural networks for learning and
memory of cognitive recognition codes. Proceedings of the 12th Annual Conference of the
Cognitive Science Society, Hillsdale, NJ: Erlbaum Associates, 1032-1034
Chikkerur, S., Serre, T., Tan C. and Poggio, T. [2009]. An integrated model of visual attention
using shape-based features. CBCL paper #278/MIT-CSAIL-TR #2009-029, Massachusetts
Institute of Technology, Cambridge, MA, June 2009
Fiser, J., Berkes, P., Orbán, G., and Lengyel, M. [2010]. Statistically optimal perception and
learning: from behavior to neural representations. Trends in Cognitive Sciences 14(3): 119-130.
Gershman, S.J., Vul, E, and Tenebaum, J. [2010]. Perceptual multistability as Markov chain
Monte Carlo inference. Advances in Neural Information Processing Systems.
Goodman, N., Tenenbaum, J. B., Feldman, J., and Griffiths, T. [2008]. A rational analysis of
rulebased concept learning. Cognitive Science, 32(1), 108-154.
Griffiths, T., and Tenenbaum, J. [2006]. Optimal predictions in everyday cognition.
Psychological Science, 17, 767-773.
Moreno-Bote, R., Knill, D. and Pouget, A. [2009]. Sampling and Optimal Cue Combination
during Bistable Perception. Computational and Systems Neuroscience Annual Meeting
(COSYNE), Salt Lake City, UT
Oliva and Torralba, [2006]. Building the Gist of a Scene: The Role of Global Image Features in
Recognition. Progress in Brain Research, 155(2), 23-36
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

O'Reilly, R.C. [1996]. The Leabra Model of Neural Interactions and Learning in the Neocortex.
Phd Thesis, Carnegie Mellon University, Pittsburgh, PA
Robertson P. and Brady, J. M. [1999] Adaptive Image Analysis for Aerial Surveillance, IEEE
Intelligent Systems, 14(3), 30-36.
Robertson, P. and Laddaga, R. [2003]. GRAVA: An Architecture Supporting Automatic Context
Transitions and its Application to Robust Computer Vision, Proceedings of the 4th International
and Interdisciplinary Conference CONTEXT 2003, Stanford, CA
Tenenbaum, J. B. [1999]. Bayesian modeling of human concept learning. In M. Kearns, S. Solla,
& D. Cohn (Eds.), Advances in neural information processing systems (Vol. 11, p. 59-65).
Cambridge, MA: MIT Press.
Tesauro, G. [1995]. Temporal Difference Learning and TD-Gammon. Communications of the
ACM 38(3).
Vul, E., Goodman, N., Griffiths, T. L., and Tenenbaum, J. [2009]. One and done? Optimal
decisions from very few samples. In Proceedings of the 31st annual conference of the cognitive
science society.

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 Biologically Inspired Cognitive Architectures 2010 23
A.V. Samsonovich et al. (Eds.)
IOS Press, 2010
© 2010 The authors and IOS Press. All rights reserved.
doi:10.3233/978-1-60750-661-4-23

SCA-Net: A Sensation-Cognition-Action
Network for Speech Processing
Michael Connolly Brady
Speech Lab, Department of Cognitive and Neural Systems, Boston University
677 Beacon Street, Boston, MA, 02215, USA
[email protected]

Abstract. Two related questions posed to biology-inspired cognitive robotics are:

(1) what form might a motor plan take in the brain, and (2) how might such a plan
be converted into coordinated motor behavior? One approach to modeling the
motor plan is to assume a stream of discrete instructions. This conceptualization
has been attractive in cognitive science because it allows theorists to discuss motor
planning (and perception) in terms of notation systems. However, many brain and
robotics researchers have come to view the idea of the discrete motor plan to be
wanting. SCA-Net provides a concrete alternative based on established brain
circuitry. The approach allows for the coding of perception and production
representations in a way that is useful to notation-oriented cognitive architectures.

Keywords. Adaptive Resonance Theory (ART), cerebellar model, motor learning

SCA-Net is based on an ART-style cortico-thalamic circuit coupled with a cerebellar

model for timing coordination. In the basic circuit, an array of input nodes sends an
auditory signal to a neural field through an adaptive filter. The field is conceptualized
as a change detector where specific changes in the auditory signal perturb the field to
shift between equilibrium states. In turn, as the field shifts between equilibrium states it
sends an output signal through an adaptive filter to the motors of an articulatory speech
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

synthesizer. Adaptive filters are updated during ‘babble learning’ where categorical
movements (shifts between distinct field equilibrium states) produce motor output that
is associated with resulting acoustic input. The circuit eventually ‘hears’ new sounds in
terms of its experience producing sounds. The basic circuit is iterated both serially and
in parallel to build a network capable of learning increasingly longer sequences.
Simulated babble learning is demonstrated using a mechanical vocal tract.

Figure 1. basic SCA-Net circuit (left), a network of these circuits controls a mechanical vocal tract (right).
Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 24 Biologically Inspired Cognitive Architectures 2010
A.V. Samsonovich et al. (Eds.)
IOS Press, 2010
© 2010 The authors and IOS Press. All rights reserved.
doi:10.3233/978-1-60750-661-4-24

A Connectionist Model of MT+/Mstd

Explains Human Heading Perception in the
Presence of Moving Objects
N. Andrew Browning
Scientific Systems Company Inc., 500W Cummings Park, Woburn, MA 01801
[email protected]

Abstract

Human heading perception is robust in the presence of moving objects, except when
the object crosses the focus of expansion (FoE). If the object approaches the observer,
heading perception is shifted towards the object FoE [1]. If the object maintains a
fixed distance, heading perception is shifted away from the object FoE [2]. This data
has resulted in the theory that differential motion operators are used to determine
heading. Primate neurophysiology, however, indicates that additive MT cells which
combine data across pools of motion sensitive cells in V1, and not differential cells, are
used to assess heading in dorsal MST [3]. The present work utilizes components of the
ViSTARS (Visually-guided Steering, Tracking, and Route Selection) Biologically
Inspired Cognitive Architecture (BICA) to analyze and reconcile the data.
ViSTARS BICA components relating to heading perception [4] were analyzed and
minimal modifications were made to provide distance dependent weighting in the
template match. The results indicate that differential motion operators are not required
to explain human heading bias, provided that a detailed model of V1-MT-MSTd is
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

used. The ViSTARS BICA indicates that heading bias in the presence of moving
objects results from the representation of heading as a template and the use of a
population codes to represent spatial properties within the system. Work is ongoing to
fully characterize the causes and nature of this bias. This work demonstrates the utility
of BICA not only in extending our computational capabilities but also in providing an
overall framework with which to understand and reconcile the data on which it is built.

Keywords. ViSTARS, heading, differential motion, MT, MST, FoE, motion

References

[1] Warren, W. H., & Saunders, J. A. (1995). Perceiving heading in the presence of moving objects.
Perception, 24(3), 315–331.
[2] Royden, C. S., & Hildreth, E. C. (1996). Human heading judgments in the presence of moving objects.
Perception and Psychophysics, 58(6), 836–856. (Elder et al. 2009;
[3] Duffy, C. J. (1998). MST neurons respond to optic flow and translational movement. Journal of
Neurophysiology, 80(4), 1816–1827.
[4] Browning, A., Grossberg, S., and Mingolla, M. (2009). A neural model of how the brain computes
heading from optic flow in realistic scenes. Cognitive Psychology, 59, 320-356.

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 Biologically Inspired Cognitive Architectures 2010 25
A.V. Samsonovich et al. (Eds.)
IOS Press, 2010
© 2010 The authors and IOS Press. All rights reserved.
doi:10.3233/978-1-60750-661-4-25

Discovering the visual patterns elicited by

human scan-path
Andrea CARBONE
Dept. of Computer and System Sciences. Sapienza, Università di Roma.

Abstract In this work we present a biologically motivated framework for the mod-
elling of the visual scene exploration preference. We aim at capturing the statistical
patterns that are elicited by the subjective visual selection and reproduce them via
a computational system. The low level visual features are encoded through the pro-
jection of the image patches on a learned basis of linear filters reproducing the typ-
ical response properties of the primary visual cortex (V1) receptive fields of mam-
mals. The resulting training set is typically high-dimensional and sparse. We ex-
ploit the sparse structure by clustering together patterns of channel activation which
are similar on the basis of a binary activation map and finally deriving a pooling
over the set of the original linear filters in terms of active (on) and non-active (off)
channels for each cluster. The system has been tested on a dataset of natural images
by comparing the fixation density maps recorded from human subjects observing
the pictures and the saliency maps computed by our system obtaining promising
results.

Keywords. Natural Image Statistics, Gaze Modeling, Sparse Coding

1. Introduction
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

The visual exploration of the world, the scan-path, is performed by human beings in a
very efficient way by programming a sequence of saccades on the visual scene in order to
project the selected spatial focus of attention onto the higher resolution spot of the retina.
As argued in [17], the fixations do not follow a uniform spatial distribution and at the
same time, several experiments have also highlighted a substantial difference between
the statistics measured at the centre of fixation and non-fixated regions [15,16]. Even if
the visual selection is likely to be influenced by a mixture of bottom-up and top-down
factors it is possible to purposely model the visual patterns fixated to learn a generative
model of the subject’s fixations. Lot of effort has been devoted to the discovery of the
sensory coding mechanism delivered by the early visual processing areas. Early works
on this subject highlighted the importance in modelling the statistical regularities of the
visual input, following the assumption that the visual system exploits those regularities
to efficiently code the visual information [2]. Nonetheless as argued in several works [5],
the purpose of the early visual processing is to produce a sparsified representation of the
visual input rather than a compression, see also [11]. In this work we are using a standard
dataset comprising 101 images depicting urban and natural scenes with corresponding
eye-tracked data from 31 subjects. Each image has associated an individual as well as a
cumulative fixation distance-map, for further reference see [8].
Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 26 A. Carbone / Discovering the Visual Patterns Elicited by Human Scan-Path

2. Linear Generative Models

A very simple but successful model assumes that any given set of natural images (or
patches) I (x, y) can be generated by a linear combination of basis vectors B :


n 
I (x, y) = Bi (x, y)si si = Wi (x, y) (x, y) (1)
i=1 x,y

where the coefficients si is the coefficient weighting the i-th basis vector, Wi is the
feature or linear filter associated to the i-th basis. In other words, the si coefficients rep-
resent the response strength of a model simple cell Wi obeying to a linear behaviour
w.r.t the (visual) input I . The estimation of the can be achieved by a statistical approach
called ICA (Independent Component Analysis). ICA is able to reconstruct the elemen-
tary original sources when constraining their distribution to be non-Gaussian and sparse.
A sparse code is a code whose response distribution is usually sharply peaked and long
tailed (super gaussian). This means that each coding unit has a low probability of being
active while most of the time will outputs values close to zero. The idea behind sparse-
ness is therefore that only a small subset of a large population of cells will be active
when presented to a specific family of visual stimuli. In this work we focus exclusively
on the characterisation on the sensory model of the simple and complex cells at the early
stages of cortical processing (V1) which exhibit a localised, oriented and bandpass be-
haviour [6,1]. The popularity of sparse coding as a statistical prior in the modelling of
has been discussed in [5] where a linear basis of the spatial receptive field of simple cells
is estimated by constraining the coefficient distribution to be maximally sparse while
minimising the reconstruction error of the training set [12]. ICA is usually solved as
a batch algorithm, each iteration updates the results considering the full set of sample
inputs. There are other approaches such as for example the Lobe Component Analysis
(LCA) [19] which instead uses an incremental-type procedure suitable for online learn-
ing with open-ended sequences of samples. The LCA approach produces as output a set
orientation-selective cells similar standard batch ICA.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

ICA learning has proved to be very efficient in modelling receptive fields tuned for
particular scenes/features or properties of the human visual processing, see [9,14]. In
some cases ICA learnt bases were used to model the feature distribution of local image
patches and compare it with its surrounding delivering an interpretation of saliency as
a local measure of self-information (or surprise) [4] or to model neural networks for
learning [18].

3. Our Approach

We suggest a combined approach mixing several aspects of computational neuroscience

and machine learning which aim at identifying those statistical patterns which trigger a
subject’s spot of attention towards specific locations. A computational model is derived
by firstly estimating a mixture of multivariate Bernoulli variables out of a binary acti-
vation map representing the visual receptive field responses to a target sequence of fixa-
tion patches. Afterwards, each component of the mixture is taken as modelling a specific

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 A. Carbone / Discovering the Visual Patterns Elicited by Human Scan-Path 27

family of receptive field responses which (statistically) share a common pattern of ac-
tivation. A dichotomized Gaussian representation lets us select for each component the
most representative sample that we elect as prototypal example of the population. The
set of prototypal samples are related to a binary representation which bring us to a nat-
ural interpretation of the patterns as a collection of on and off maps Fig. 1. The on-off
channels map respectively those channels whose response are expected to lie on the tail
(a relatively high value) or the peak (close to zero).

Figure 1. From the scanpath to the Mixture of Bernoulli and finally its corresponding mapping of on and off
filters.

3.1. Receptive field estimation

The Independent Subspace Analysis (ISA) [7] is based on the maximisation of indepen-
dence between linear subspaces. The principal difference
  is that the basis are grouped to-
gether into subspaces S(k) of finite size: I (x, y) = k i∈S(k) Bi (x, y)si . The group-
ing is useful to model how the complex cells pool the inputs from the simple cells, thus
characterising a non-linear feature detector. The pooling is implemented
 as the square-
root of sum of the squares of the subspace coefficients: uk = 2
i∈Sk si , where the si
are computed as in Eq. 1, see Fig. 2. The uk are also called energy detectors. ISA is thus
configured as a first linear filtering followed by an energy pooling stage. The result is a
set of linear filters whose pooled response reproduce closely the response properties of
complex cells in V1 (independence of the response magnitude w.r.t. the phase of the sig-
nal). The ISA basis has been computed on a set of 60000 (24 × 24) pixel image patches
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

randomly selected from the image collection.

Figure 2. Non linear energy detector as square root of the sum of squares of the linear responses of the feature
detector belonging to the subspace.

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 28 A. Carbone / Discovering the Visual Patterns Elicited by Human Scan-Path

3.2. Scan-path projection on ISA basis and Mixture of Bernoulli estimation

We define the target scan-path SP as a set of fixation patches filtered out from a gaze-
tracked sequence. The matrix F contains on each row the uj coefficients. The size of
F depends on the number N of fixations and on the geometry of the ISA basis: the
total number of basis D, the subspace dimension d and M = D d , the dimension if the
pooled invariant feature space. The input dataset will be denoted by F ∈ RM ×N . In our
experiments we computed a complete linear basis of dimension D = 576 for a 24 × 24
image space. The subspace size was set to d = 8, thus the pooled feature space has
dimension M = 72, see Fig. 3. Our goal is to capture the pattern of activation of the
invariant feature channels by transforming the original continuous-valued model vector
F into its thresholded (binarized) version F01 , thus discarding activation values which
are under a certain threshold. The binarisation may seem radical however, in this context,
we are focused in modeling the overall coefficient responses as a family of binary spiking
cells, each of them firing only when the input visual stimuli matches the tuning properties
of the individual cell.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

Figure 3. The response population to a single patch, the ISA linear coefficients and the pooled subspace
non-linear coefficients output.

A threshold τ for each of the M ISA invariant subspaces is defined as follows. Let
ukj be the value of the j-th patch projected on the k-th non-linear subspace and sij with
i ∈ Sk be the linear coefficients belonging to the k-th subspace. Then τ ∈ RM is an
optimal threshold, such that ukj = 1 or ukj = 0 depending on whether ukj is above
or below τk , if the overall activity sparseness of the binary code is maximised while
minimising the average reconstruction error induced by choosing only the sij coefficients
with i ∈ Sk related to the k-th subspaces which are set to “on” after thresholding (i.e.
set to 1). The activity sparseness is a measure of population sparseness, a property of
the response distribution of all the cells to a single stimuli. It is defined as “how few
cells are active at any one time” [20]. The reconstruction error instead accounts for the
loss induced by the new encoding. Sparsification and reconstruction are two contrasting
measures in sensory coding, an high sparseness could decrease the reconstruction power

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 A. Carbone / Discovering the Visual Patterns Elicited by Human Scan-Path 29

of a code. A perfect reconstruction may require too many components to be active at the
same time. In other words we aim at finding that value for τ minimising the following
functional:

    
τ ∗ = arg min Sp(F̂) j + α Err(F, F̂) j , j = 1...N (2)
τ


F̂ being the thresholded version of F. The operator . in this context means the
average over all the patches of the function between the brackets. The constant α weights
the contribution of the mean squared error with respect to the sparseness-related term.
This expression has a similar structure to the one presented in [13]; intuitively the first
term is to preserve the original coefficient sparseness the ICA receptive field, the second
term instead wants to preserve the capacity of the code to reconstruct the original patches
appearance. The two terms are defined for the i-th row in S as:

 
M  2
Sp(F̂)j = #{ŝij | i ∈ Sk , ûkj = 1} , Err(F, F̂)j = Ij − ûkj sij Bi
k=1 i∈Sk
(3)
where ûkj is the binarized version of the ukj . Therefore in Eq. (3) only the ICA/ISA
coefficients corresponding to the 1’s actually contribute to the linear reconstruction. A
standard gradient descent algorithm is used to estimate the best τ ∗ as Eq. (2). The initial
value of τ is set to the standard deviation of each uk . After thresholding we obtain F01 , a
binary matrix of the same size as F which is the spike train activation map corresponding
to the original train of scan-path patches. We model the F01 as samples of a mixture of
multivariate Bernoulli (MoB). The parameters of the mixture have been estimated via
expectation-maximisation (EM) algorithm for Bernoulli mixtures as reported in [3]. In
Fig. 4 the plot of F01 with an example of the clustering induced by the estimated Mixture
of multivariate Bernoulli distribution.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

Figure 4. F01 , the thresholded F responses and clusters emerging after the Mixture of Bernoulli estimation.

4. The computational model

We link the binary spike population estimated mixture to its corresponding dichotomized
Gaussian (DG) [10] generative counterpart. The properties of such decomposition are

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 30 A. Carbone / Discovering the Visual Patterns Elicited by Human Scan-Path

exploited by finding out the most probable subspace (in the DG space). We show how
such subspace encodes a representative activity pattern of the receptive field responses
elicited by the target visual input.

4.1. The dichotomized Gaussian model (DG)

A population of N neurons which can fire or not when presented with an input signal,
can be in one of the 2N states. The DG distribution is a statistical tool described in [10]
used to generate spike trains with given first and second order statistics by considering
dichotomized samples of a multivariate Gaussian distribution N (γ, Λ). Samples from the
DG reproduce a population of binary patterns whose statistics match the desired second
order correlations. Given the i-th cell, a binary sample Xi is obtained by firstly drawing
a sample Ui from the multivariate Gaussian U ∼ N (γ, Λ) and then thresholding to get
a 0 or a 1: Xi = 1 iff Ui > 0. For each of the M components the set F01 m
∈ F01
is the subset of the binarized scan-path samples that are more likely to belong to the m-
th component. From each F01 m
we estimate the corresponding DG model N (γm , Λm ).
At this point, we have at our disposal a generative model of the binarized input dataset
representing the full scan-path. It inherits the structure from the previously estimated
MoB, resulting in a weighted sum of M multivariate DG distributions.

4.2. From the mixture DG model to the on-off filters

The idea is to make use of the MoB and its DG model decomposition to implement a
filtering cascade that will output a high saliency value corresponding to local images
patches that have a pattern of responses which is close to the target scan-path when pro-
jected on the receptive fields. Intuitively, in the dichotomized representation we can iden-
tify the most probable binary pattern that can occur by considering the sign of γm . In
fact, in a N dimensional space there are 2N subspaces centred on the origin depending
on the sign chosen on the n-th individual dimension. Each of these subspaces by defini-
tion is mapped to a specific binary pattern. We want to identify the most probable binary
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

pattern that can be sampled from N (γm , Λm ). The subspace where the mean of the DG
falls it is of course the subspace whose cumulative distribution is greater than the others.
Therefore we take sign(γm ) as the prototypical binary activation pattern which approx-
imates the m-th component. The negative sign will be mapped to a 0 and the positive
to a 1. The ISA-subspace feature channels corresponding to the ones and the zeroes will
model respectively the on and off linear filtering cascades for each DG component. The
on and off responses to a single patch input are given by:


M D 
M D
on = πi uj bij of f = πi uj (1 − bij ) (4)
i=1 j=1 i=1 j=1

where πi are the mixture coefficients of the i-th component, uj are the subspace
coefficients for the input patch and the bij are 1 or 0 depending on the previous interpre-
tation of the mean of the DG model of the i-th component. The on-off maps computed
on every (24 × 24) patch of the input image are then merged together to get the final
saliency map SM.

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 A. Carbone / Discovering the Visual Patterns Elicited by Human Scan-Path 31

5. Results and Conclusion

We use the metric suggested by Tatler et al in [17] to evaluate and compare the perfor-
mances of computational models for the prediction of visual saliency. The ROC curve
displays the sensitivity of a binary classifier when varying a decision threshold. The
curve plots the false positive rate (FPR) vs. the true positive rate (TPR) for each value
of the threshold. The source of the ground-truth data is given by the fixation density
maps corresponding to each test image. The maps are binarized by setting all the values
above a threshold τf to 1 , thus splitting the image into fixated (ones) and non-fixated
(zeroes) regions. The computation of the ROC curve proceeds as follows: the saliency
maps produced by our system are normalised in the range (0, 1) and then progressively
thresholded with τs delivering a binary map identifying predicted fixated (ones) and non-
fixated (zeroes) regions. For each of the threshold values τs applied the density map and
the resulting saliency map are compared. Depending on the corresponding binary labels
for the ground truth and the saliency map the TP and FP rates are finally computed. In
Fig. 5(a) a test image from the category “buildings” of the dataset it is shown, the com-
puted saliency map and the cumulative fixation distance-map of all the subjects. The
brighter regions (red in the colour version) indicate high saliency values, in Fig. 5(b)
the ROC curve representing the performance of the system on the overall category. The
results are promising, considering that there is no explicit model for the temporal aspects
of the subject’s selection and that the system is learning perceptual patterns that often
are repeated across the whole image, thus assigning a high saliency value regardless of
the image coordinates, whereas the subject’s fixations originating that pattern are likely
to be localised in specific locations of the image driven by a contextual visual salience.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

Figure 5. (top) An input test image, its computed saliency map and the reference fixation distance map.
(bottom) ROC curve: AUC: 0.684, Original Fixation Density Map thresholded at 90% of the original range.

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 32 A. Carbone / Discovering the Visual Patterns Elicited by Human Scan-Path

References

[1] R Arkin. Integrating behavioral, perceptual, and world knowledge in reactive navigation. Robotics and
Autonomous Systems, Jan 1990.
[2] H Barlow. The exploitation of regularities in the environment by the brain. Behav Brain Sci, 24(4):602–
7; discussion 652–71, Aug 2001.
[3] Christopher M Bishop. Pattern recognition and machine learning (information science and statistics).
2006.
[4] Neil D B Bruce and John K Tsotsos. Saliency, attention, and visual search: An information theoretic
approach. J. Vis., 9(3):1–24, 2009.
[5] D. J Field. What is the goal of sensory coding? Neural Computation, 6:559–601, Jan 1994.
[6] D Hubel and T Wiesel. Receptive fields and functional architecture of monkey striate cortex. The
Journal of Physiology, 195:215–243, Jan 1968.
[7] A Hyvarinen and U Koster. Complex cell pooling and the statistics of natural images. Network: Com-
putation in Neural Systems, 18(2):81–100, 2007.
[8] G Kootstra, A Nederveen, and B de Boer. Paying attention to symmetry. Proceedings of the British
Machine Vision Conference (BMVC), pages 1115–1125, 2008.
[9] U Koster, JT Lindgren, M Gutmann, and A Hyvarinen. Learning natural image structure with a hori-
zontal product model. Independent Component Analysis and Signal Separation, pages 507–514, 2009.
[10] JH Macke, P Berens, AS Ecker, AS Tolias, and M Bethge. Generating spike trains with specified
correlation coefficients. Neural Computation, 21(2):397–423, 2009.
[11] B. A Olshausen and D. J Field. Natural image statistics and efficient coding. Network: Computation in
Neural Systems, 7(2):333–339, 1996.
[12] B. A Olshausen and D. J Field. Wavelet-like receptive fields emerge from a network that learns sparse
codes for natural images. Nature, Jan 1996.
[13] B. A Olshausen and D. J Field. Sparse coding of sensory inputs. Current Opinion in Neurobiology,
14(4):481–487, 2004.
[14] Hyun Jin Park and Te-Won Lee. Capturing nonlinear dependencies in natural images using ica and
mixture of laplacian distribution. Neurocomputing, 69(13-15):1513–1528, 2006.
[15] P Reinagel and AM Zador. Natural scene statistics at the centre of gaze. Network: Computation in
Neural Systems, 10(4):341–350, 1999.
[16] C A Rothkopf and D H Ballard. Image statistics at the point of gaze during human navigation. Visual
neuroscience, 26(01):81–92, 2009.
[17] BW Tatler, RJ Baddeley, and ID Gilchrist. Visual correlates of fixation selection: Effects of scale and
time. Vision Research, 45(5):643–659, 2005.
[18] J Weng and M Luciw. Dually optimal neuronal layers: Lobe component analysis. IEEE Trans. Au-
tonomous Mental Development, Jan 2009.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

[19] J Weng and MD Luciw. Optimal in-place self-organization for cortical development: Limited cells,
sparse coding and cortical topography. Proc. IEEE 5th International Conf. on Development and Learn-
ing (ICDL 2006), 2006.
[20] B Willmore and DJ Tolhurst. Characterizing the sparseness of neural codes. Network: Computation in
Neural Systems, 12(3):255–270, 2001.

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 Biologically Inspired Cognitive Architectures 2010 33
A.V. Samsonovich et al. (Eds.)
IOS Press, 2010
© 2010 The authors and IOS Press. All rights reserved.
doi:10.3233/978-1-60750-661-4-33

An Architecture For Humanoid Robot

Expressing Emotions And Personality
Antonio CHELLAa and Rosario SORBELLOa,1
and Giorgio VASSALLOa and Giovanni PILATO b
a
Dipartimento di Ingegneria Informatica, Università degli Studi di Palermo,
Viale delle Scienze, Palermo,Italy
b
ICAR, Istituto di Calcolo e Reti ad Alte Prestazioni, Italian National Research
Council,Viale delle Scienze, Palermo,Italy

Abstract. In this paper we illustrate the cognitive architecture of a humanoid robot

based on the proposed paradigm of Latent Semantic Analysis (LSA). The LSA
approach allows the creation and the use of a data driven high-dimensional
conceptual space. This paradigm is a step towards the simulation of an emotional
behavior of a robot interacting with humans. The Architecture is organized in three
main areas: Sub-conceptual, Emotional and Behavioral. The first area processes
perceptual data coming from the sensors. The second area is the “conceptual space
of emotional states” which constitutes the sub-symbolic representation of emotions.
The last area activates a latent semantic behavior related to the humanoid
emotional state. The robot generates its overall behavior also taking into account
its “personality”. To validate the system, we implemented the system on a
Aldebaran NAO humanoid robot.

Keywords. Humanoid Robot, Emotions, Personality, Latent Semantic Analysis.

Introduction
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

The ability to recognize and understand emotions plays an important role for a
natural and spontaneous human robot communication.
In last years, many efforts have been made towards the experiments in the field of
human robot social interaction [3],[9]. Research by Breazel [8] has shown an humanoid
robot’s emotion model as important basis in the social interaction with humans. Arkin
et al. [5],[6] presents an emotion model for the architecture of an entertainment robot.
Liu et al. [10] describes an emotion model of 3D virtual character. Miwa et al. [7] has
illustrated a mechanism for humanoid robot WE-4RII to express in a natural way
human-like emotions.
The interaction of entertainment robots with humans and the environment should
not be mechanical and deterministic in order to avoid a predictable and trivial behavior.
To reach this goal we have modeled a Cognitive Architecture that drives the behavior
of a humanoid robot taking into account its own “personality”, the context in which it
is immersed, the present stimulus and the most recent behaviors executed by the robot.

1
Corresponding Author.
Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 34 A. Chella et al. / An Architecture for Humanoid Robot Expressing Emotions and Personality

This cognitive architecture of a humanoid robot is based on the paradigm of Latent

Semantic Analysis [3]. This paradigm is a step towards the simulation of an emotional
behavior of a robot interacting with humans [1],[2],[4].
The presented approach integrates traditional knowledge representation and
associative capabilities provided by geometric and sub symbolic information modeling.
The remainder of the paper is organized as follows: in section 1 the creation of the
emotional space of the robot is illustrated; in section 2 the cognitive architecture of the
humanoid robot expressing emotions and personality is described; finally in last section
3 the experimental results conducted with the humanoid Robot Nao Aldebaran are
shown.

1. The Emotional Space Creation

The architecture of the robot is based on the creation of a probabilistic emotional

conceptual space automatically induced from data. The methodology is inspired to the
Latent Semantic Analysis (LSA) technique [12].
The approach is based on the application of the Truncated Singular Value
Decomposition (TSVD), preceded and followed by specific pre- and post-processing
phases that consent to give a probabilistic interpretation of the induced vector space.
The procedure has been widely explained in [2],[11],[13].
Let us consider a corpus made of N text chunks, each expressing an emotion, and
let M be the number of words in the dataset. The goal is to realize a mapping between
the M words and the N text chunks verbally describing emotions into a continuous
vector space S, where each word, as well as each emotion is associated to a emotional
knoxel in S.
In analogy with knoxels in conceptual spaces, an emotional knoxel is a vector in the
probabilistic emotional space; from the conceptual point of view, it is the
epistemologically basic element at the considered level of analysis.
Let A be the M×N matrix whose (i,j)-th entry is the square root of the sample
probability of the i-th word belonging to the j-th text chunk. The Singular Value
Decomposition of the matrix A is performed, so that A is decomposed in the product
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

of: a column-orthonormal M×N matrix U, a column-orthonormal N×N matrix V and a

N×N diagonal matrix , whose elements are called singular values of A.

A USV T (1)

Let us suppose that A’s singular values are ranked in decreasing order. Let R be a
positive integer with R<N, and let UR be the M×R matrix obtained from U by
suppressing the last NR columns, R the matrix obtained from  by suppressing the
last NR rows and the last NR columns and VR be the N×R matrix obtained from V
by suppressing the last NR columns. Then

AR UR SR VR T (2)

is a M×N matrix of rank R. The matrix AR is then post-processed in order to obtain a

new matrix  in this manner:

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 A. Chella et al. / An Architecture for Humanoid Robot Expressing Emotions and Personality 35

­0 if >aR @ij < 0

°
>\ @ij
°
® >aR @ij
otherwise
°
¦ >aR @ij
2
°
¯ i, j:> a R @ ij ! 0
(3)

It can be shown that the illustrated procedure can be seen as a statistical estimation
process, being  the probability amplitude estimated starting from the sample
probability amplitude A.

2. The Cognitive Architecture

The architecture of the presented system is inspired to the approach illustrated in

[14], [15], and it is organized, as shown in figure 1, in three main areas: the sub-
conceptual area, the emotional area, and the behavioural area.

2.1. The Sub-Conceptual Area

The sub-conceptual area is aimed at receiving stimuli from the different modalities
of the robot and controlling at a low level also its actuators. Two main modules
compose this area:
x The MotionModule controls the actuators of the robot in order to perform the
movement requested by the Behavioral Area.
x The PerceptualModule processes the raw data coming from robot sensors in
order to obtain information about the environment, and the external stimuli.

The perceptual information sensed by the processing of the raw sensor data is
associated to their English description: as an example, if the system recognizes a “red
hammer”, it will be associated to the sentence “I can see a red hammer”. The use of a
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

verbal description of the information retrieved from the environment allows their
mapping into the emotional area in an easy way. These natural language descriptions
associated to each modality will be the input of the conceptual area.

2.2. The Emotional Area

The emotional area makes the robot able to find emotional analogies between the
current status and the previous knowledge stored on the robot using the semantic space
of emotional states.
The associative area is built up in order to reflect and encode not only emotions and
objects that provoke emotions, but also the personality of the robot. This is true in a
twofold manner:
x documents used to induce the associative space characterize its dimensions and
concepts organization beneath the space;
x personality is also encoded as a set of knoxels that play the role of “attractors” of
perceived objects beneath the space.

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
 36 A. Chella et al. / An Architecture for Humanoid Robot Expressing Emotions and Personality

A corpus of documents dealing with emotional states has been chosen in order to
infer the space. Emotional states have been coded as “emotional knoxels” in this space
using verbal description of situations capable to evoke feelings and reactions.
Environmental incoming stimuli are also encoded using natural language sentences.
We have selected the following emotional expressions: sadness, fear, anger, joy,
surprise, love and a neutral state. A corpus of 1000 documents, equally distributed
among the seven states, including other documents characterizing the personality of the
robot, has been built. This set of documents represents the emotional knowledge base
of the robot. The excerpts have been organized in homogeneous paragraphs both for
text length and emotion.
A matrix has been organized where the 6 emotional states and the neutral state have
also been coded according to the procedure illustrated in the previous section.
Each document has been processed in order to remove all “stopwords” i.e. words
that do not carry semantic information. A 87x1000 terms documents matrix A has been
created where M=80+7 is the number of words plus the emotional states and N=1000
is the number of excerpts. The generic entry aij of the matrix is the square root of the
sample probability of the i-th word belonging to the j-th document. The TSVD
technique, with K=150, has been applied to A in order to obtain its best probability
amplitude approximation < . This process leads to the construction of a K=150
dimensional conceptual space of emotions S . The axes of S represent the
“fundamental” emotional concepts automatically induced by TSVD procedure arising
from the data. In the obtained space S , a subset of ni documents for each emotional
state corresponding to one of the six “basic emotion” Ei has been projected in S using
the folding-in technique, i.e. each excerpt is coded as the sum of the vectors
representing the terms composing it. As a result, the j-th excerpt belonging to the
subset corresponding to the emotional state Ei is represented in S by an associated
vector em ji and the emotional state Ei is represented by the set of vectors

^ em i : j 1,2,K n i .
 j
` (4)

The personality of the robot is encoded as a set of knoxels derived from documents
Copyright © 2010. IOS Press, Incorporated. All rights reserved.

describing the fundamental personality characteristics of the robot. As an example, if

the robot has a “shy” personality, a set of documents dealing with shyness, bashfulness,
diffidence, sheepishness, reserve, reservedness, introversion, reticence, timidity, and so
on are used to construct a cloud of “personality knoxels” that represent attraction points
for the external stimuli that provoke emotions in the robot. Therefore the i-th excerpt
belonging to the subset corresponding to the personality characteristic is mapped as a
vector pi in S .
The inputs from the sense channels are coded in natural language words or sentences
describing them and projected in the conceptual space using the folding-in technique.
These vectors, representative of the inputs from the channels, are merged together as a
weighted sum in a single vector f(t) that synthesizes the inputs stimuli from
environment at instant t:
The input stimulus is therefore “biased” through the computation of the contribution
of personality attractors in the space as the weighted sum of the “personality knoxels”.
pi representing the personality of the robot, Each personality knoxel can be weighted
with coefficients Di (with 0 d DI d1) in order to fine tune the personality influence upon
the robot’s behaviors.

Biologically Inspired Cognitive Architectures 2010 : Proceedings of the First Annual Meeting of the BICA Society, edited by A. V.
Random documents with unrelated
content Scribd suggests to you:
If you were fortunate enough to secure a Senate hearing for your
bill the chairman of your committee appointed a sub-committee of
about three; in the House, the entire committee might attend the
hearing. A day was set and you began preparing your ammunition;
the opposition was allowed an equal amount of time to the second.
After the hearing a vote was taken. If they were against it, they
killed it then and there; if they recommended it, it came up before
the full committee and, if then approved, went to the Senate or
House for debate on the floor.
To the frantic, worried, harassed, driven Congressmen of 1931 the
announcement of a birth control bill was like a message from Mars,
only less interesting and more remote. The mind of each Senator
resembled a telephone switchboard with his wary secretary as the
operator. All the wires were tied up with foreign debts,
unemployment relief, reparations, moratoriums, sales taxes,
prohibition, budgets and bonuses, war in Manchuria, peace
conferences, disarmament, and the tariff—issues of vital concern to
themselves for which they needed every vote; and their principal
endeavor was not to cause conflict or get themselves disliked. What
chance had we to plug in?
When the vigilant secretary found we were not direct constituents,
we were told the Senator was busy—in conference, in committee,
meeting an arriving delegation. Would we come back later,
tomorrow, next week? Always we came back promptly and on the
dot. For months it was almost impossible to see any of them. Often
as many as forty calls were made, and if we succeeded in getting
two interviews, we considered that a good day’s work. When finally
we did reach them, few of the younger, still fewer of the older,
Senators knew what we were talking about. When we were able to
make this clear, young and old alike, just as in the state legislatures,
were full of fears—fear of prejudices, fear of cloakroom joshings,
mainly fear of Catholic opposition.
Though Senator Norris had approved the repeal bill, he believed that
ours had a better chance of passing because antagonism to the
former was even greater than in 1926. He himself had Muscle Shoals
and the Lame Duck Amendment on his hands and several more pet
projects to boot, and suggested we get somebody to introduce the
bill who would not be up for re-election. Our choice fell on Senator
Frederick Huntington Gillett of Massachusetts, for years Speaker of
the House, and now about to retire. He was a gentleman born, gray-
haired, typically New England, without children or any particular
philosophy regarding birth control. Our Southern helpers, notably
Mrs. J. B. Vandeveer, were persistent and determined. They would
not be put off with polite, routine dismissals, but asked point-blank,
“Will you introduce this bill for us?” Senator Gillett, recognizing their
earnestness, agreed. But we heard no more of it.
When I returned at the next session of the same Congress someone
remarked, “Aren’t you lucky to have had your bill introduced?”
“What?” I stared with wide-open eyes.
“Yes, Senator Gillett remembered it a few days before the session
closed.”
I called on him at once. “Where’s our bill gone?”
It had gone nowhere. “We’ll just send it around to the Judiciary
Committee,” said the Senator. “Norris is Chairman and he’s friendly.
He’ll pick out a good sub-committee for you.”
We gathered our witnesses together the night before the hearing,
which was to be February 13th, and asked, “What do you want to
say? How long do you want in which to say it?” We had eight people
to testify in the space of two hours; moments had to be carefully
parceled out to each. We were permitted to deduct ten from our
allotment the first day to be used the following one for a rebuttal.
William E. Borah of Idaho and Sam G. Bratton of New Mexico had
been assigned to us with Senator Gillett, but Borah did not appear.
The audience, mostly women, crowded the committee room,
imposing with marble pillars, glossy mahogany, gleaming windows.
Dr. John Whitridge Williams, obstetrician in chief of Johns Hopkins,
summed up the medical evidence for birth control. “A doctor who
has this information (prevention of conception) and does not give it
cannot help feeling he is taking a responsibility for the lives and
welfare of large numbers of people.” The Reverend Charles Francis
Potter, founder of the Humanist Society of New York, discussed the
moral phase. “The bird of war is not the eagle but the stork.”
Professor Roswell H. Johnson, then at the University of Pittsburgh,
stressed eugenics. “Most intelligent, well-informed people ... are so
determined in this (spacing children) that no laws yet devised
succeed in forcing a natural family, which is about eighteen children,
upon them.” Rabbi Sidney Goldstein dealt with religious aspects.
“The population is not made up of those who are born but is made
up of those who survive.” Professor of Sociology Henry Pratt Fairchild
spoke from the economic point of view. “We human individuals
cannot break laws of nature. We can, however, choose which of her
laws we see fit to obey.” Mrs. Douglas Moffatt announced that the
twenty-seven hundred members of the New York City Junior League
were overwhelmingly in favor of the bill.
The next morning the opposition began by trying to prove that we
who advocated birth control, a Russian innovation, were seeking to
pull down motherhood and the family as had been done in Russia.
The Honorable Mary T. Norton, Representative from New Jersey,
made the astounding assertion that the happiest family was the big
one, and that a large percentage of the great men and women of
this country were born poor; this was a blessing since it fired them
with ambition. And she mentioned Abraham Lincoln, whose birthday
had been but two days before. I was particularly outraged by
hearing statements from other witnesses that the American
Federation of Labor was against us, that the American Medical
Association was antagonistic, and that the Methodist and other
churches were going to help defeat our bill. Speaker after speaker
representing Catholic organizations repeatedly hurled such dramatic
tirades as, “I ask you, gentlemen, in the name of the twenty million
Catholic citizens of the country, to whose deep religious convictions
these vices are abhorrent, and of all those to whom the virtue of a
mother or a daughter is sacred, to report unfavorably on this
diabolical and damnable bill!”
It was difficult to gauge the impression that was being made; you
could only sense that the response was one of feeling. These
dogmatists, harking back to the Dark Ages, summoned to their aid
the same arguments that had been used to hinder every advance in
our civilization—that it was against nature, against God, against the
Bible, against the country’s best interests, and against morality. Even
though you proved your case by statistics and reason and every
known device of the human mind, the opponents parroted the line of
attack over and over again; in the end you realized that the appeal
to intelligence was futile.
On occasions like this the inward fury that possessed me warmed
from coldness to white heat; it did not produce oratory, but it
enabled me to move others. The way to meet the opposition was to
keep emotions in hand and, at the same time, without stumbling or
fumbling, to let them go. Every word I said was calculated and
thought through, not in advance, but as it came along. I did not
react this way often, but I did that day.
When my ten minutes for rebuttal came, I knew that emotional
speed was required. Nevertheless, I first knocked down their false
assertions: that the birth control movement had originated in this
country during 1914, long before anyone had ever heard of
Bolshevism; that the objections of the American Federation of Labor
had referred to the repeal bill of 1925, quite different from the
doctors’ bill now under discussion; that the American Medical
Association had taken no stand, but two of its most important
branches, the Neurological and Woman’s Medical, had gone on
record in our favor; that Dr. C. I. Wilson of the Methodist Board had
denied his church was opposed, and, in fact, its ministers had
worked unofficially for us. “When someone says that the happiest
families are the largest ones, and that the world’s great leaders have
been of large families, I would like to call to your attention that the
great leader of Christianity, Jesus Christ himself, was said to be an
only child.” Here the Catholics crossed themselves and muttered,
‘Blasphemy!’
“These opponents have had the laws with them, the wealth, the
press, and yet they have come today to say they are afraid of the
morals of their people if they have knowledge, if they do not
continue to be kept in fear and ignorance. Then I say their moral
teachings are not very deep. Mr. Chairman, we say that we want
children conceived in love, born of parents’ conscious desire, and
born into the world with sound bodies and sound minds.”
The two Senators sat there in silence. The bill was killed, due to the
adverse vote of Senator Borah—who had not attended the hearings.
The next year, 1932, Senator Gillett was gone and a substitute had
to be found. Believing the first woman Senator would be on the side
of her sex, we asked Mrs. Hattie Caraway to introduce the bill. She
said she herself was interested in the subject, but her secretary
would not let her touch it.
Ordinarily Congressmen paid little attention to abstract arguments,
logic, or the humanitarian needs of outsiders. But they could be
reached through their constituents. One way of doing this was to get
women “back home” to help themselves directly by writing letters.
This required money. We sought it from a foundation which donated
ten thousand dollars earmarked for this special purpose. To the still
continuing stream of letters from mothers, requesting as always
contraceptive advice, my reply went, “I would gladly give you the
information you ask for if the law permitted. Your Congressman now
has the opportunity to vote on this bill. Send him a letter telling how
many children you have living, how many babies dead, how many
abortions, what wages your husband receives, everything you have
told me,” and I enclosed an envelope, stamped and addressed to
their respective Congressmen.
While walking one day through the tunnel which connected the
House with the Senate, I stopped to ask a man my direction. He
said, “I’m going your way. Come along and I’ll show you.”
We fell into conversation. He informed me he was a Senator, and
asked what I was doing.
“I’m working on the birth control bill.”
“That’s funny. I’ve just had a letter from a woman five miles from
where I’ve lived most of my life. Listen to this.”
And he took it out of his pocket and read the history of the woman’s
abortions and operations. “I’ve never heard anything quite so awful,
and at the bottom she says, ‘You can help me by getting this law
changed, and Mrs. Sanger, who has the information, will send it to
me if you get the law changed.’”
These letters brought fine results. Through them Senator Henry D.
Hatfield of West Virginia was persuaded to introduce the bill. At the
hearing he described how as physician and surgeon and governor of
his state he had seen the free mating of the unfit, and had forced
through a sterilization law. We produced our usual array of experts,
and the opposition produced Dr. Howard Atwood Kelly, a famous
gynecologist in his day at Johns Hopkins, but now Professor
Emeritus and very old, who rambled discursively on morals; his was
a state of mind if not of reason. Dr. John A. Ryan, a member of the
National Catholic Welfare Conference, chose economics for his
discussion. Neither spoke on his own subject, but selected
something on which he was not an authority.
The bill was killed in committee, and the one introduced by
Representative Frank Hancock of North Carolina in the House got
into the wrong committee so nothing happened.
Before you had seen it, the Congress of the United States loomed
impressively in your consciousness; you had a feeling, “This is the
greatest country in the world, this is its Government, I helped to
send these men here.” Then you watched Congress at work, listened
to it, and were disillusioned. A few years of sitting in the gallery and
looking down gave you less respect for the quality of our
representatives, less faith in legislative action, and you wondered
whether those who had already abandoned hope of obtaining relief
in this way and resorted to direct action had not, perhaps, the right
idea.
The same arguments went on from year to year. A certain amount of
publicity was secured, a certain number were educated. Some of our
followers, in face of the evidence to the contrary, still were confident
that if the Catholics understood our bill they would not obstruct it.
They said Representative Arthur D. Healey of Massachusetts, a
member of the Judiciary Committee, although a Catholic was so
liberal that if he could once be made to see the reasons back of it he
would cease being openly hostile, and it might even get out of
committee. Accordingly, I went to his office; we talked at length, and
again got nowhere. As I was leaving this father of four said, in order
to explain himself, “You see, Mrs. Sanger, I’m just one of those
unusual men who are very fond of children.” I was inwardly
convulsed at the thought that he considered himself unusual and
that we were all a lot of Herods trying to do away with babies.
At first it seemed that I was to have greater success as the result of
my interview with Dr. Joseph J. Mundell, Professor of Obstetrics at
Georgetown University, who advised the Catholic Welfare Conference
on all their medical legislation. In a private session I conceded some
things in the bill; Dr. Mundell gave up certain others. The
compromise apparently suited everybody.
In 1934 identical bills were introduced in Senate and House, the
latter by Representative Walter M. Pierce, Democrat, who as
Governor of Oregon had burned his political bridges by vetoing a bill
which permitted parochial schools. Since he had nothing to lose, he
did not have to play politics.
Hatton W. Summers of Texas was chairman of the hearing. Our side
led off, again specialists in each line covering the vital points. Rabbi
Edward L. Israel of Baltimore made an impassioned plea. “And I say,
gentlemen, if this thing we are now advocating is not morally right,
let us stop being hypocrites and, in its place, put a law on our
statute books that will drive contraceptive devices out of your homes
and mine.”
Here John C. Lehr of Michigan, sitting back in his chair with thumbs
hitched in his suspenders, declared pompously, “As a member of this
Committee I want to go on record there have never been any
contraceptives used in my home. I have six children, too.”
Malcolm C. Tarver of Georgia interrupted, “You don’t mean any
member of Congress has used anything of that kind, do you?” His
surprise was obviously genuine.
The proponents of our bill, even elderly women, had stood while
delivering their testimony. But when Father Charles E. Coughlin
entered, cheeks very pink over his black collar, a chair was placed for
him, because as a representative of the Church he would not stand
before a representative body of the State. He began talking at
random, “I have not heard one word of the testimony these ladies
and gentlemen have produced, and my remarks are not addressed
to them now, because I can easily handle them over the radio
Sunday after Sunday.... You, gentlemen, you are married men, all of
you, and you know more about it than I will ever know.” Here he
arched his eyebrows into a leer. “The Chairman, I understand, is a
bachelor like myself.... We know how these contraceptives are
bootlegged in the corner drug stores surrounding our high schools.
Why are they around the high schools? To teach them how to
fornicate and not get caught. All this bill means is ‘How to commit
adultery and not get caught.’”
Some of our sympathizers walked out of the room. Two
Congressmen left the table. But we were a polite, well-behaved
group that shrank from scenes, and, though furious and indignant,
we allowed him to conclude his half-hour of grossness.
I could hardly believe my ears when Dr. Mundell, who shortly before
had helped us formulate a bill which he said was satisfactory to him,
rose and deliberately betrayed us by stating there was no need for
legislation whatsoever, because a recent scientific work—by which he
meant Rhythm—had shown that fertility in women could be
reckoned with almost mathematical precision.
In the rebuttal Dr. Prentiss Willson testified that the theory of the
cycle of sterility had no medical standing. Then came my turn. I had
in my pocket a copy of Rhythm, and quoted from it. Under the
heading of procreation it asked whether married people were obliged
to bring into the world all the children they could, and then made
answer:
Far from being an obligation, such a course may be utterly indefensible. Broadly
speaking, married couples have not the right to bring into the world children
whom they are unable to support, for they would thereby inflict a grievous
damage upon society.
I told the committee that apparently the only distinction in the pros
and cons of the birth control question was that the method we
advocated was a scientific one under the supervision of doctors; that
of the Catholics had not been proved scientifically and was open to
any boy or girl who could read the English language.
Nevertheless, the bill again died in committee.
The Senate hearings on the bill, introduced by our old friend Daniel
O. Hastings of Delaware, did not come until March. We presented
our advocates, among them a miner’s wife from West Virginia, the
native state of two members of the committee, Hatfield and Nealy.
She was a perfect illustration of the type which most needed birth
control. When she had finished a Catholic woman asked her, “Which
of your nine children would you rather see dead?”
“Oh, I don’t want to see any of them dead. I love them all; but I
don’t love those I haven’t had.”
Her reply was just right; it could not have been better.
Vito Silecchia, my former coal and ice vendor from Fourteenth
Street, also made his way to Washington and told his simple story.
His wife had come to me when pregnant with her fourth child, and I
had said I could do nothing for her until she had had her baby. Now,
many years afterwards, she had no more than the four. Vito
reasoned his case as a man, “I am a Catholic myself. The Catholics
say we should have much children. I say different. I say it is not
good to have too many children. You can’t take care of them.” He
ended by describing the mother of six who lived next door to him. “I
told her, ‘I will take you to a place. It is a wonderful place.’ She does
not know the English language. Therefore, she has never come up
to see Mrs. Sanger, but she will—but she will!”
For the first time the Senate sub-committee reported out the bill and
it was put on the unanimous consent calendar. The last day of the
session came, June 13th. Over two hundred were ahead of it, but
there was always hope. One after another they were hurried through
and then, miracle of miracles, ours passed with no voice raised
against it. The next one came up, was also converted into law,
another up for discussion, tabled. Twenty minutes went by. Suddenly
Senator Pat McCarran from Reno, Nevada, famous divorce lawyer
though an outstanding Catholic, came rushing in from the cloak
room and asked for unanimous consent to recall our bill. As a matter
of senatorial courtesy Senator Hastings granted his request; had he
not done so Senator McCarran would have objected to every bill he
introduced thereafter. It was summarily referred back to the
committee and there died.
In 1935 we took the fatal step of having it voted on early in the
session and it was promptly killed. The whole year’s labor was lost.
The following winter, when I was in India, Percy Gassaway of
Oklahoma introduced a bill in the House, Royal S. Copeland of New
York, in the Senate, by request; neither one reached a hearing.
Another line of attack on the Comstock law was to try for a liberal
interpretation through the courts. Among the products shown at the
Zurich Conference in 1930 had been a Japanese pessary. Pursuing
the clinic policy of testing every new contraceptive that appeared, I
ordered some of these from a Tokyo physician. When notified by the
Customs that they had been barred entrance and destroyed, we sent
for another shipment addressed to Dr. Stone in the hope that it
would then be delivered to a physician. But this also was refused,
and accordingly we brought suit in her name.
After pending two years the case finally came up for trial before
Judge Grover Moscowitz of the Federal District Court of Southern
New York. Morris Ernst conducted our claim brilliantly, and January
6, 1936, Judge Moscowitz decided in our favor—the wording of the
statute seemed to forbid the importation of any article for preventing
conception, but he believed that the statute should be construed
more reasonably. The Government at once appealed and the case
was argued in the Circuit Court of Appeals before Judges Augustus
N. Hand, Learned Hand, and Thomas Swan, whose unanimous
decisions were rarely reversed in the Supreme Court.
In the fall of 1936, while I was in Washington getting the Federal bill
started again in advance of Congress’ meeting, news came that the
three judges had upheld the Moscowitz decision and had added that
a doctor was entitled not only to bring articles into this country but,
more important, to send them through the mails, and, finally, to use
them for the patient’s general well-being—which, for twenty years,
had been the object of my earnest endeavor.
The Government still had the right to appeal inside of ninety days.
Therefore, I was not unduly jubilant. We had had so many seeming
victories that melted away afterwards.
But long before the period of grace had expired, Attorney General
Cummings announced to the press that the Government would
accept the decision as law, and, with commendable consistency, the
Secretary of the Treasury sent word to the Customs at once that our
shipments should be admitted. It is really a relief to be able to say
something good about the Government.
In the face of the court decision there was little point at this time in
continuing the Federal campaign. The money for closing it up came
through a most unexpected and affecting channel. About a year
after I had seen Viola Kaufman at the California Conference in 1931,
I received a letter from her asking me please to write out the form in
which I would like any money left so that she could designate it in
her will. I took her clear, concise note to my attorney who suggested
that, since organizations were many and might go out of existence
at any moment, it would be wiser to have the bequest in my name
to be dispensed for any purpose within the movement I saw fit. I
answered her to this effect and she replied, “I am now passing over
to you in my will whatever I possess.”
I considered that the only courteous thing to do was to have Anna
Lifshiz, who was living in Los Angeles, go to see Miss Kaufman. The
address was in the Mexican district, in the poorest, most dilapidated,
run-down section. In patched clothes she came to the door of her
house, in which there was hardly any furniture. She was formal and
rather cold.
Anna merely explained the reason for her call was that she knew
Miss Kaufman as one of our subscribers. She wrote me, “That poor
creature hasn’t money enough to keep body and soul together.”
Two years went by. I was in Washington, preparing to start for
Boston for a meeting when a messenger boy delivered a telegram
from the director of the General Hospital at Memphis, Tennessee,
requesting me to come at once; Viola Kaufman was dangerously ill
with pneumonia and asking for me. I looked up trains; it would take
forty-eight hours, and so I put in a long-distance call to the director,
who told me she had died during the night.
“What was she doing in Memphis?”
“We don’t know. The Salvation Army brought her in to us. She has
only a little cash tied up in a handkerchief. We can’t do anything
without you because you’re the beneficiary.”
The undertaker also wanted an order from me, and, since her
executor, an officer of a bank in Los Angeles, had gone on a fishing
trip, I arranged the details for her cremation. She had ordered that
her remains be sent to me and when they arrived the clinic staff
came up to Willow Lake and we held a little memorial service of
gratitude and respect, spreading the ashes over the rock garden.
To everybody’s astonishment Viola Kaufman had about thirty
thousand dollars in Los Angeles realty. But it took a year and a half
to settle the estate and by this time everything was at the lowest
ebb of the depression. We received approximately twelve thousand
dollars. I have never looked at the obituaries for the last twenty
years without hoping to read that someone has willed a million
dollars for birth control, but the only legacy ever bequeathed us was
that saved from the meager earnings of this schoolteacher, Viola
Kaufman, who herself lived in poverty.
With this money we wrote finis to the Federal legislation. Of the old
organization all that was left in Washington was a secretary to read
the Congressional Record daily—a watchdog to report any bills
proposed which would make it necessary for us to jump into action
to combat them.
Six years of this work had cost one hundred and fifty thousand
dollars. It had also meant strain and worry beyond anything I had
ever attempted—never being able to detach myself from it whether
Congress was in session or not, always on the alert to discover any
new person elected who might be favorably disposed. Now and
again it had been discouraging; you could exert yourself to the
utmost with pleasure if it were a matter of convincing a person and
watching his mind being pried open, but here, over and over again,
you saw this same conviction, yet he reverted to the same fears and
refrained from doing anything.
However, the process of enlightening legislators had also unclosed
the eyes of an enormous number of organizations. First to approve
publicly had been the National Council of Jewish Women. Eventually
more than a thousand clubs—civic, political, religious, and social,
including the General Federation of Women’s Clubs, the Y.W.C.A.,
local Junior Leagues—in all representing between twelve and
thirteen million members—had given their endorsement. And, more
important than anything else, the public had been educated
persistently, consistently away from casual and precarious
contraceptive advice into the qualified hands of the medical
profession.
Dr. Dickinson had been appearing regularly at American Medical
Association meetings, keeping the question constantly alive. But not
until Dr. Prentiss Willson had formed a national body of doctors in
1935 to carry on legislative work had there been any action. One
had stirred up; the other organized.
I was at Willow Lake one June morning of 1937 when I saw spread
across the newspaper in double column the glad tidings: the
Committee on Contraception of the American Medical Association
had informed the convention that physicians had the legal right to
give contraceptives, and it recommended that standards be
investigated and technique be taught in medical schools.
In my excitement I actually fell downstairs. To me this was really a
greater victory than the Moscowitz decision. Here was the
culmination of unremitting labor ever since my return from Europe in
1915, the gratification of seeing a dream come true.
These specific achievements are significant because they open the
way to a broader field of attainment and to research which can
immeasurably improve methods now known, making possible the
spread of birth control into the forlorn, overpopulated places of the
earth, and permitting science eventually to determine the
potentialities of a posterity conceived and born of conscious love.
 Chapter Thirty-five

A PAST WHICH IS GONE FOREVER

P arenthood remains unquestionably the most serious of all human

relationships, the most far-reaching in its power for good or for
evil, and withal the most delicately complex. I always tried to secure
my sons’ confidence by being honest with them, treating them as
though they had intelligence, and expecting them to use it. For the
sake of companionship it was essential to be honest, no matter what
the cost. Fortunately, the younger generation is not crumpled up
when sharply confronted with the truth. They have cut through the
regard to their feelings until they can say extraordinarily blunt things
to each other and yet not be hurt. And with this they have invented
a new language; they can “take it.”
Many times I could have forced my opinion on the boys and saved
Fern perhaps some bitter disappointments—“Let me do it. I’ll
manage all this. Let me know when you need anything.” But,
instead, I merely stated my attitude and said, “Here are the two
alternatives. You want this; I think the other is better. Neither of us
can tell which is right. If you choose your own way I’ll help you as
long as you do it well, providing you stop as soon as you know it is
wrong and go back and pick up the other. If experience teaches you
a greater wisdom, you can call it square.”
At Peddie Institute, Stuart was paying more attention to sports than
studies. It was easy for him to be an athlete. But he also had a
logical mind and a quick ability for co-ordinating hand and brain.
When he was ready for college he entered Sheffield Scientific School
of Yale University. His imagination was soon captured by archaeology
and medicine, but his course was already set.
Meanwhile Grant, who had been inclined to hero-worship his older
brother, had also gone to Peddie. His athletics left little opportunity
for bringing out his artistic talents, and he agreed to take his last
two years at Westminster School in Simsbury, Connecticut, where he
was encouraged to develop along his own lines. In his sophomore
year at Princeton, he still had no idea of what he wanted to do with
his life. Although he had a leaning towards diplomacy, which would
include training in law, I explained to him that, since the family had
no political influence, it might lead to being a small politician.
And so I made out a list of as many occupations known to man as I
could think of, and sent them to him, telling him to mark off with a
blue pencil those which he was perfectly sure did not appeal to him,
and check with red those for which he felt some predilection. Out
immediately went piano-mover, waiter, floorwalker, bank manager,
bookkeeper, and some fifty others.
Six months later, I returned him the red-checked list for further
perusal. Now his preferences were much more definite. Research,
journalism, editorial work, diplomacy were again red, but almost
everything else marked headed him for a scientific career.
The decision made, Grant began his pre-medical course.
After Stuart graduated from Yale he moved downtown to Wall Street
and continued in a broker’s office all during the depression. But, in
this money making atmosphere, his attitude was changing. He had
concluded that serving humanity was a higher fulfillment than
profiting at humanity’s expense, and medicine seemed the career
which he also liked best. Having found out, he had the courage to
start back at the beginning to accomplish it. We made a compact for
him to go as far as he could and test whether his interest kept up.
First he had to acquire sufficient chemistry and biology, going to
Columbia University in the daytime for the former, to New York
University in the evening for the latter, preparing his lessons until
three in the morning.
The next year he passed his entrance examinations.
Following the legislative near-victory in the winter of 1934, I
resolved to go to Russia to see for myself what was happening in the
greatest social experiment of our age. With keen anticipation I
looked forward to discovering whether the Marxian philosophy,
dramatized and realized and based on an economic ideology, did not
have to accept some of the philosophy of Malthus.
Grant, then about to enter his final year at Cornell Medical School,
was eager to investigate the progress of medicine in the Soviet
Union, and made up his mind to come along. I was taking also my
secretary, Florence Rose, efficient, competent in any capacity,
whether field organizing or in the office. Though but recently
enlisted in the movement, she had come more with the attitude of
the early days, not for what she could get out of it, but for what she
could give to its furtherance. Her talents and enthusiasm, when
added to her cheerfulness, made her a rare combination; always
gleeful and bubbling with fun, she carried out nearly everything in
that spirit.
Mrs. Ethel Clyde, an officer of the Federal legislative organization,
was to be the fourth of our little group within a large group. When
zeal for the “new civilization” in Russia had been at its height she
had relinquished her expensive Park Avenue apartment for a smaller
one on a side street, and contributed the difference in rent to sundry
leftist causes and birth control.
At the last moment it seemed we might not be able to go. For some
years Stuart had had a bad sinus condition, and hardly had he
matriculated at Cornell in the fall of 1933 when he had been struck
by a squash racket, fracturing the bone over his eye. That winter he
had been operated on nine times. A week before I was due to sail
this doctor advised that he have an exploratory operation. I rushed
up from Washington, where the legislative work for that session was
just being wound up, and would have abandoned the Russian
expedition had not the operation apparently been entirely successful.
Stuart insisted that I go. Since he was in no danger I continued with
my plans.
It was not feasible to travel in Russia except in a party under official
guidance. Three people I knew who had gone by themselves
described how train after train had passed them, boat after boat had
steamed down the Volga with no accommodations available.
Therefore, we chose the non-partisan Second Russian Seminar.
Shortly prior to leaving I spent an evening with Maurice Hindus, Will
Durant, John Kingsbury, and Drs. Hannah and Abraham Stone, all of
whom had been to Russia the previous year. Maurice Hindus had
returned impersonal and still unprejudiced, Will Durant utterly
antagonistic, John Kingsbury full of fervor, and both Stones warmly
disposed. They had all been in Moscow, practically at the same time,
for approximately the same number of days, and all had received
utterly dissimilar impressions. Even pictures that Will Durant had
taken were not the same as those of John Kingsbury or Dr. Stone,
snapped from almost identical places, thus showing me how wide
might be the variety of responses, depending on the individual bias.
I expected to keep my eyes open, to think independently, to ask
questions, and compare. I was going to use as much sanity and
fairness as I possessed, and not be swept emotionally into any
current of opinion.
Billy Barber was the manager of the Seminar, and I did not envy him
his job. There were many complaints and stupid remarks and much
faultfinding. Most of the party were going merely to be able to say
those things were true which they had previously said were true. I
asked one woman who went on every sight-seeing expedition but
never got out of the bus, “Why did you come?”
“Oh, just to wipe Russia off my list.”
Edward Alsworth Ross was among the leaders. He was the only
person who had been there under the former regime some twenty
years earlier, and had an authoritative basis of contrast between the
old and the new; we all rather sat at his feet. He was a typical
professor, wore enormously high, stiff collars, played checkers with
anybody who would indulge him, and was upset when he failed to
win. His personality was impressive, literally so because wherever
you looked you spied him. One of the funniest sights was to see this
Nordic giant, six feet four, walking with short dark Florence Rose,
five feet two, each jollying the other.
We scooted through England across to Copenhagen, about which I
recall very little. I was always trying to learn what advance the
women’s movement had made, but somebody was always trying to
tell me how marvelous the city was. Remembering Ellen Key, I
reached Scandinavia with great hopes for Feminism. But the women
who were considered the most intelligent were complacently resting
on their laurels. The older ones still reigned supreme and believed
that, because they had won their battles of twenty-five years ago,
there was nothing left to fight for. The younger group found it hard
to rise above the inertia of this overwhelming prestige. Since
population was not a problem in Scandinavia, they were interested
chiefly in eugenics, and had almost forgotten the aspect of individual
suffering.
At Oslo a number of us went on pilgrimage to the grave of Ibsen. As
I stood there in silent tribute I had the feeling he had understood
women and the ties they had been loosening. To my mind Nora
never went back to the “doll’s house”; her evolution was too
complete. Or, if she did return, she entered by another door.
Mr. Barber had arranged to feed his hundred and six charges at the
last Finnish railroad station. There was a particular exhilaration
about the prospect of that meal, because it was to be our final one
before crossing into “famine-stricken” Russia. We arrived at ten in
the morning, all of us hungry. As we filed into the station our eyes
met the most gorgeous panorama—long tables beautifully laid out
with delicious meats, fish, breads, compotes.
While we paused, debating which of these delicacies to taste first,
there came a stampede of fifty other Americans, a tourist group led
by Sherwood Eddy. Never had I seen such an exhibition. The men,
unshaven, hatless, coatless, pushed and shoved around, in front of,
and almost on top of the tables. The best we could do was find
comfortable seats from which we could have a good view of the riot.
The meal prepared by the railroad with such courtesy for our party
was demolished by another.
Barber and Eddy eventually discovered it was all a mistake. The train
carrying the Eddy-ites had failed to stop at the town where their
repast had been awaiting them, and naturally they supposed this
breakfast was theirs.
At Leningrad we were met by buses and driven through streets that
swarmed with imperturbable, peasant-like people. The upper parts
of their Mongolian-shaped heads all looked exactly the same. I
noticed how immaculate they were. Faces, necks, hands, were white
as white and displayed a cleanliness simply marvelous when you
took into consideration the difficulty of securing soap and water. Very
few were old; many were children apparently between the ages of
two to twelve. But in the expressions of all I glimpsed a sadness.
The former capital was depressing and down at heels, shabby and in
need of painting. Yet it was beyond comparison in its spacious
dignity; the architectural design of the houses could not be hidden.
My high-ceilinged room at the Astoria was luxurious with alcove bed,
bath room, and large marble tub, which, although cracked and
spotted with rust, nevertheless evidenced the days of splendor when
the hotel had been frequented by the aristocracy of the Old Regime.
From my window I could see the cobbled square. It was eight
o’clock and the city was awakening. I watched the passing show:
heavy wagons were drawn by a single and often most decrepit horse
with what seemed a dark brown rainbow, arched and graceful, over
his neck; queues formed in front of little stands that served rations
of beer or bottled soda water; some women, the varying colors in
their shawls making bright splotches, swept the car tracks with birch
switches or pushed empty carts on their way to market, others
carried hods of cement up the ladders to the masons on the new
buildings being erected everywhere. Usually the men were doing the
skilled work, and women, hardy and robust, with strong legs, bare
feet, sunburned faces, were kept at the laborious, monotonous,
physical labor until such time as they could qualify as expert
artisans.
The Communists’ apartments were much better, lighter, airier,
cleaner, more modern than those for non-party members. When we
asked why, in an equalitarian state, one section should be thus
privileged, we were answered, “It was they who made all this
possible. Why should they not have the best? What you bourgeois
give to your capitalists, we give to our Communists.”
We asked Tanya, our guide, if she were a Communist, and she
replied, “Oh, no. That’s too hard.” Ordinary citizens might be
excused for a mistake or even a crime, but party members could
have no human frailties. They were exiled or perhaps shot for
cheating, stealing, deceiving, exploiting, taking money under false
pretenses, or many things which average people could do and be
punished with fines alone.
Although the cost of the trip itself was relatively low, whatever we
bought in Russia was excessively high owing to the peculiar situation
of the ruble. In the first place, there was no ruble; it existed only in
theory. Second, every foreigner was supposed to deal exclusively
with the Torgsin stores through which the Government had cleverly
contrived to come by a hoard of foreign currency by charging
seventy-eight cents in our money for each ruble instead of its actual
value of five cents. For example, the price of a stamp on a letter to
the United States, which was two and a half rubles, amounted to
two dollars.
Mrs. Clyde, who leaned sympathetically towards Communism, said to
one of our young men, “Let me get you a little present.”
“Not here,” he said. “It’ll be too expensive.”
“Oh, yes,” she insisted. “What would you like?”
“Well—a bar of almond chocolate, then.”
She had to pay ten American dollars for that ten-cent bar of
chocolate. Her Communism melted slightly.
Ultimately, we solved the ruble problem. One morning a boy who
had been loitering around the Astoria asked Grant, “Would you like
me to take you through the city?”
Grant prudently inquired, “How much?” It appeared that the boy
merely desired an opportunity to perfect his English; he had plenty
of rubles, which he was glad to dispose of at the rate of fifty for a
dollar. Russians could obtain none but the cheapest commodities on
their tickets; if they wanted luxuries such as good shirts, leather or
rubber boots, and other articles sold only at Torgsin, they were
obliged to surrender some treasured gold piece or use foreign
money.
With an ample supply of rubles I sent long, elaborate cables to
Stuart to cheer him up. He must have thought an excessive maternal
solicitude was getting the better of my economic judgment. But, as a
matter of fact, one of twenty words was costing me less than
twenty-five cents.
Dr. Nadina Kavanoky, who had been interested in birth control in the
United States, had given me a letter to her father, Dr. Reinstein, once
a dentist in Rochester, New York, now in Stalin’s close confidence.
He came to see me about eleven-thirty one night, the Russian calling
hour, and we talked until three in the morning. When he wanted to
know my “impressions of Russia,” I said promptly, “It seems to me
your policy of overcharging us is a mistake; for the sake of a few
dollars you are creating ill will, just as the French have done. In our
own Seminar we have twenty librarians and perhaps double that
number of schoolteachers and students, many of whom have gone
without other vacations to come here. They have a unique
opportunity to influence people; everybody will ask them when they
get back, ‘Did you like Russia?’ You are trying to build up a favorable
public opinion abroad, and these people are the best mediums for
that purpose. If they are pleased they will fight for you and break
down prejudice.”
But he was not convinced, and, evoking the specter of the Tsarist
debt to America, he replied, “We’ll bleed you, we’ll milk you, we’ll
get every dollar out of you we can. America demands her pound of
flesh and this is how we’ll pay you.”
The occasions for receiving “pleasant impressions” were offered by
vigorous tours to points of interest. We were given a choice of hard
buses or harder ones, all, in my experience, springless and clattering
noisily over the cobble-paved streets. After a few bumps we usually
hit the roof and came down with headaches. Our poor little guides
had to screech with full lung power to be heard over the incessant
rattling.
One morning when driving back from sight-seeing, the motor gasped
and collapsed on a slight hill. Passengers volunteered helpful
suggestions—“Put it in low. Put it in neutral. Push this. Pull that.”
The driver moved gears forward and backward and then looked
around at us in perplexity, “I did, but it won’t work.” We waited and
waited and waited and waited. Somebody ran a mile to telephone
that we were stranded and needed another bus. Meanwhile,
everything we wanted to see was closing, and we had already
learned that whatever you missed in Russia was always the most
worth while. In fact, it seemed they had visiting hours timed to end
five minutes before you got there. Several other buses came along
and stopped. Their drivers got out, poked their heads under the
hood, began taking things apart, strewing bolts this way and nuts
that. Then they, too, became discouraged, and, leaving increased
confusion, climbed on their chariots again and went on.
Finally some bright young man discovered we were out of gas.
As we crossed the huge square in front of the hotel, I saw directly
ahead of us an enormous pile of bricks with wide spaces on both
sides. Closer and closer we came. “When will the driver turn?” I
asked myself. But he never did; we went right over the top and the
bricks slipped out from under. That was the Russian system. You
could not go round an obstacle; you must go over it.
Enlarged portraits of Lenin and Stalin were in all public buildings.
Their statues were everywhere, in every square, on every corner. A
major industry of Russia seemed to be to find new poses for Stalin—
standing up, lying down, writing, reading. Often just his head,
definitely recognizable in spite of the predominance of red, was
designed in flower beds. One of the most delicate attentions was to
give him a different colored necktie on different days; the plants
were kept in pots to make this charming gesture possible.
After the Revolution when peace had come, connoisseurs from
various countries had been invited to examine the recovered statues,
rugs, tapestries, and objets d’art stolen from the palaces and
churches. One by one the priceless paintings were displayed,
specialists rendered their opinions, commercial dealers furnished
appraisals, stenographers took down every word. The same was
done with the lapis lazuli tables, the snuff boxes, the court jewels.
The interesting part of the new arrangement was that the
interpretation was entirely Marxian. Pictures, instead of being hung
according to the orthodox history of art, were fitted into the
Industrial Revolution. A certain Madonna was not admired for its
qualities of color or form, or as a thing of beauty in itself; the guide
explained to you that it was created at such and such a time when
the Church was trying to get a hold over the people, when artists
were starving and had to look for their means of livelihood to the
patronage of the Church.
Later, in the Kremlin at Moscow we saw fantastic and incredible
riches, jeweled saddles, a whole set of harness studded with
turquoise, a huge casket cloth embroidered with thousands of
pearls. In order to place the period of the latter I asked Tanya where
it had come from. She replied in her precise English, “You see, it is
for to cover the dead. You see, in Russia there was such a custom.
When they died they put them in the ground. It was such a custom,
you see, to cover them with cloths.”
She spoke of the Tsarist Regime as though it had been centuries
ago.
One of the pictures was a Christ removed from the cross and lying
on the ground. Tanya said, “People used to come here, and they
even kissed it!” This she uttered in the tone of scorn of a very
youthful generation shocked and horrified at the ancient traditions.
“Our hope is in the young people,” she said frequently.
“But how old are you?”
“Oh, I’m thirty-two,” as though she were doddering.
Grant and I were once walking by a group of children when a small
boy pointed at us and remarked, “Ah, there go some of the dying
race.” To them all Amerikanski were capitalists.
The Marxian ideology had been applied to every phase of life. H.G.,
accompanied by Gyp, his biologist son, had flown over from London.
Since he wanted an opportunity to go around alone, he rather
resented being so closely guarded and courteously guided. After
talking with Stalin he had come to the conclusion that the Dictator
had no understanding of economics. He was somewhat annoyed at
the constant interpretation of everything in terms of politics, and of
having Marx stuffed down his throat at every turn.
At the schools you might ask what kind of mathematics they taught.
“Marx.”
“And what system of engineering?”
“Marx.”
No matter what the question, the answer was Marx.
The Anti-Religious Museum, once a cathedral, was directly across
from the Astoria. Each half-hour little girls, who seemed hardly more
than ten or twelve, their sleeves hanging down over their finger tips,
with great dignity conducted excursions of peasants through. Their
lecture started with the fundamental principle that the earth was
round. A bas-relief of the world was underneath the huge pendulum
which hung from the dome. If you stood there long enough you saw
it swing from one point to a further one. They were trying to show
that it was within man’s power to make his own heaven.
Here were kept the relics of the churches, the icons laden with silver
and gold wrung from the poor peasants in the past. Actual concrete
things were reduced to their simplest terms on large poster-type
murals which depicted stories, a necessary practice since the
muzhiks were so generally illiterate. In one a kulak was coming to
the priest with a sick child in his arms, asking for prayers to cure its
illness. The priest, fat and clad in rich robes, shook his head, saying,
“You must bring money for the saint. The saint will not cure your
child unless her arms are covered with silver.” But the kulak had only
his farm. “Mortgage it and get the money,” the priest ordered. Soon
the kulak returned with silver, and the mural showed how now the
saint’s arm was almost hidden. But still the child remained sick. “The
saint’s halo is bare,” said the priest. At last the whole figure was
silvered, but the baby died just the same.
Opposite this mural was pictured the Soviet way. The father carried
the baby to the hospital, where nurses with gauze across their
mouths took it preciously, bathed it carefully, laid it in bed. The
entire sterilizing process was illustrated—the doctor in white gown
and cap, scrubbing and washing each hand five minutes as marked
by a clock. Finally you saw the child, healthy and well, jumping into
its mother’s arms.
The people stood there looking, their imaginations fired. They said,
“This is what is happening to us.”
Most particularly I wanted to investigate what had been done for
women and children in Russia, to learn whether they had been given
the rights and liberties due them in any humanitarian civilization.
Grant, Rose, as she was known to me, and I went one day to the
Institute for the Protection of Motherhood and Childhood, a vast
establishment stretching over several miles, with model clinics,
nurseries, milk centers, and educational laboratories. I was
overwhelmed in contemplating the undertaking. There was no doubt
that the Government was exerting itself strenuously to teach the
rudiments of hygiene to an enormous population that had previously
known nothing of it. Russia was also aiming to free women from the
two bonds that enslaved them most—the nursery and the kitchen.
All over the country were crèches connected with the places they
worked.
Children were the priceless possessions of Russia. Their time was
planned for them from birth to the age of sixteen, when they were
paid to go to college, if they so desired. No longer were they a drain
or burden to their families. Not only were teachers or parents
forbidden to inflict corporal punishment, but children might even
report their parents for being vindictive, ill-humored, disorderly, and
in many cases they did so.
In one divorce dispute as to custody of the offspring, the father
argued that the mother was bad. The Judge asked, “Of what does
her badness consist?”
“She is nervous and loses her temper.”
The Judge agreed she was not fit for motherhood.
Furthermore, Russia was investing in future generations by building
a healthy race. If there were any scarcity of milk the children were
supplied first, the hospitals second, members of the Communist
party third, industrial groups fourth, professional classes fifth, and
old people over fifty had to scrape along on what they could get,
unless they were parents of Communists or closely associated with
them.
I was eager also to find out what had been done about the study
carried on by Professor Tushnov, of the Institute for Experimental
Medicine, on so-called spermatoxin, a substance which, it had been
rumored, produced temporary sterility in women. I made an
appointment with him, but a shock awaited me. He had tried out his
spermatoxin on thirty women, twenty-two of whom had been made
immune for from four to five months, but now all laboratory workers
had been taken from pure research and set at utilitarian tasks such
as the practical effects of various vocations on women’s health.
Nothing concerning immunization to conception could be published
in Soviet Russia, no information could be given out under penalty of
arrest, and, moreover, nothing could appear in a foreign paper which
had not already been printed in Russia.
Intourist, the Government tourist bureau, and Voks, the All-Union
Society for Cultural Relations with Foreign Countries, had asked me
when I had first arrived whom I wished to see and where I wished
to go, and had offered to call up people on my list and arrange for
visits, a service which had saved me much trouble and expense. In
spite of this co-operative attitude, I was suspicious that much was
being hidden from us. Before I had left America I had heard I could
see only what Russia presented for window-dressing, and with this
in mind I was on the alert.
Both Grant and I wondered how the hospitals built under the Tsars
compared with recent ones. When I asked to be taken to a certain
one, I was assured it was too far away, and anyhow it was being
renovated; there was nobody there. I said to myself, “Aha! here is
one of the forbidden sights. Whoever heard of a hospital equipped
to handle thousands of patients being utterly empty? They are not
going to let us see this because it might speak in favor of the old in
contrast to the new.”
Politely but firmly I insisted. Again I was told there were so many
other interesting things it would be a pity to waste my time going to
see it. I found it difficult to say anything further without giving
offense. Then Grant encountered a young American nurse from the
Presbyterian Hospital in New York who spoke Russian; she also
wanted to visit hospitals. We engaged a car of our own and drove a
good fifteen miles out of the city over horrible roads, winding and
dusty and badly paved, and even pushing on as rapidly as we could
we did not get there until late in the afternoon. To our dismay we
discovered not a patient, doctor, or nurse in the place, only
plasterers, painters, carpenters, and cleaners, pulling down and
refurbishing. We had lost half a day and were a little ashamed of our
lack of faith.
The night came to take the train for Moscow. Nobody called “All
aboard!” in Russia. Trains went right off underneath you when you
had one foot on the platform and one on the step. They just moved
and moved fast. But we clambered on and soon the leather seats
were made into our beds; they were so slippery that we kept falling
out.
Once at Moscow, we who were coming second-class, according to
Marxian procedure, received the worst rooms at the hotel; those
who traveled third had the best. I could not applaud the one
selected for me. It was directly over the laundry, and the smells of
cooking and suds floated through the window. I refused to stay and
was accommodated on the top floor where the servants had once
lived.
Moscow was as different from Leningrad as New York City from a
sleepy Pennsylvania town. The people walked more quickly and
seemed to be going somewhere, not simply wandering listlessly.
Bedlam existed at the hotels, but by now we were beginning to learn
that the Russians were so concerned with their own efficiency that
they had no time to do anything. To be in a hurry merely
complicated matters. I could wait, but for energetic Rose it was
torture. To all specific requests they replied, “It cannot be. It cannot
be.” She had her own methods of coping with this, saying she did
not wish to hear the word, “impossible”; she had no intention of
asking the impossible. Then when they procrastinated with, “a little
later,” she countered, “In America we say, ‘now!’”
Her triumph over dilatoriness came on Health Day. Since health was
almost a god in Russia, all activities ceased on that occasion and the
populace of Moscow came together on Red Square. The spectacle
was to start at two in the afternoon, but before it was light you
could hear the songs of men, women, and children moving towards
their appointed stations.
Out of our party only thirty were privileged to receive tickets, and
their names were posted. Mrs. Clyde and I were on the list, but not
Grant or Rose. The previous day the numbers were cut to twenty;
that morning there were but sixteen, and feeling ran high. “Why
haven’t I a ticket?”
Fortunately for me I had been invited to lunch by Ambassador
William C. Bullitt, who entertained lavishly and was helpful to
traveling Americans. When I had met him back in New England, I
had never thought of him as an ambassador, nor as a man skilled in
dealing with the great problems that required strategy, diplomacy,
political sagacity, and a prime knowledge of economics and history. I
considered him rather as amusing, an excellent dinner host, and one
to whom you could go when in difficulty, sure that he would get you
out. Perhaps this was what Russia wanted at that time more than
anything else. No doubt he was then somewhat disappointed at the
turn relations between Russia and the United States had taken.
Russians on the whole admired him; they had not forgotten that,
although he was not counted a proletarian or in the category of Jack
Reed, he had lifted the cudgels for them in the early days when
friends were needed.
The Ambassador’s little daughter Ann, aged ten, officiated at the
head of the table, apparently enjoying herself. The house in which
they were living while the new Embassy was being built had an
architecture quite befitting what I imagined the style of Russia
should be—a bit of the Kremlin, a bit of a mosque, and a bit of an
Indian palace.
On the way to the Square after luncheon a wave of people surged
between the rest of the diplomatic party and myself, but I kept
saying “diplomatique,” and was bowed through to the grandstand.
Meanwhile Rose had been devoting her whole attention to tickets—
and there were no tickets. The lucky holders lined up and filed off
under a leader. Rose, the ever resourceful, donned a red bandanna
and said to the “forgotten men” in the party, “We’ll make our own
battalion.” She handed out slips of paper about the size of the tickets
and then started, Grant and the Harvard professors following her
through the blare of music and the tramping troops and the
pageantry of blue trunks and white shirts, orange trunks and cerise
shirts.
Whenever anyone stopped Rose she pointed ahead and repeated my
open sesame, “diplomatique,” and they let her by until she reached
the last barrier. There the guard was suspicious of her password and
challenged her. Then she spied another group coming up, dashed
over to the leader, and exclaimed, “Quick, please explain that our
interpreter has gone on with our tickets!”
The woman looked unbelieving, but still others arrived at that
moment, and the Russian system collapsed under pressure. In they
all piled, and Rose turned to her unknown benefactress, “You don’t
know how grateful I am to you for getting us in.”
The reply was, “You don’t know how grateful I am to you for getting
us in! I’m a tourist too, and we have no tickets either.”
Nobody seeing Moscow that day could have thought it a somber
place. It was alive with song, happy faces, bright attire. The parade
of a hundred thousand or more was one of the most marvelous
spectacles for color, form, cadence, geometrical precision that I had
ever seen human beings accomplish. Men and women were
representing all sorts of games and sports—swimming, shooting,
tennis, flying. There was nothing tawdry. Each company held aloft
beautifully designed placards as it passed Stalin, who stood on top
of Lenin’s tomb. The Dictator looked much like his pictures, with his
heavy black mustache resembling the wings of a bird of prey.
All day long and everywhere you heard the Internationale, over and
over and over again. Each band struck up as it approached the Tomb
and kept playing as it swung on. Always the stirring song from those
coming up, those far away—overtones, undertones, thrilling,
insistent, now loud in your ears, now dimly echoing in the distance,
a rhythmic motif symbolizing the onward march of Young Russia.
 Chapter Thirty-six

FAITH IS A FINE INVENTION

“There is a great difference between traveling to see countries

and to see people.”
JEAN-JACQUES ROUSSEAU

“T ovarish —— wishes to see you,” came a call from the hotel

desk. For a moment I could not place the name, and the face
had changed so completely that I could but faintly trace a
resemblance to the boy I had seen before. He reminded me I had
known him in Seattle as one who had assisted in getting up birth
control meetings. When the Wobblies were being arrested in the
United States he had hired out as a stoker on a boat, and gradually
made his way to Russia, where he thought he could help to usher in
the new society.
Here was one person who had not had the best of the bargain. He
was shabbily dressed and looked dilapidated, evidently having seen
hard times, and had a beaten expression in his eyes. Yet,
disillusioned as he was, he had not come to complain. Since it was
four in the afternoon, the lunch hour in Russia, I asked him to join
me in the dining room, conducted like a large commons. The waiters
seemed disgruntled, unhappy, inept and knew very little about
service; they glanced scornfully at the man who sat down beside
me. The one lively note was the orchestra, which threw itself into
marches and wild and spirited Caucasian or Slavic folk dances while
we ate.
My guest said this was the best meal he had had since leaving
America. “Why don’t you come back?” I asked.
“I couldn’t get in.”
“Would you if you could?”
“Just give me a chance!”
I suppose it was inevitable that in such a social upheaval many
suffered. I called upon Dr. Peter Tutyshkin, who had tried to attend
our 1925 Conference in New York, but had arrived too late. As was
the case with most professional men of his years, he had been of
the old aristocracy. He and his wife and two daughters, both
physicians, had owned a beautiful home. Now the thousands of
volumes of what had formerly comprised his fine medical and
scientific library had been taken away, and he and his wife slept and
ate in the room which had contained them. He was margined and
rationed to the last degree, and I could feel his humiliation at having
so little food that he could not offer us a cup of tea.
While we were in Moscow, the Eddy party and the select six whom
Louis Fischer was piloting, crossed our path. Fischer, a Russian living
in Moscow and writing for the Nation, published in the United States,
invited Grant and me to go along with them to meet the Secretary of
the Commissariat of Public Health, Dr. Kaminsky. We went up a wide
open stairway like that of a courthouse and into a spacious room
with high windows running from floor to ceiling in French fashion
and a huge banquet table laden with the invariable afternoon tea.
Dr. Kaminsky addressed us. “Our worst heritage from the Old
Regime was in the field of medicine. The main task before us is to
unite science and practice. Our medicine is a form of social
insurance, our medical policy based on prevention. We are not
interested in profit, only service.”
Welcome to our website – the perfect destination for book lovers and
knowledge seekers. We believe that every book holds a new world,
offering opportunities for learning, discovery, and personal growth.
That’s why we are dedicated to bringing you a diverse collection of
books, ranging from classic literature and specialized publications to
self-development guides and children's books.

More than just a book-buying platform, we strive to be a bridge

connecting you with timeless cultural and intellectual values. With an
elegant, user-friendly interface and a smart search system, you can
quickly find the books that best suit your interests. Additionally,
our special promotions and home delivery services help you save time
and fully enjoy the joy of reading.

Join us on a journey of knowledge exploration, passion nurturing, and

personal growth every day!

ebookbell.com