Professional Documents
Culture Documents
William R. Uttal - The Neuron and The Mind - Microneuronal Theory and Practice in Cognitive Neuroscience-Routledge (2016)
William R. Uttal - The Neuron and The Mind - Microneuronal Theory and Practice in Cognitive Neuroscience-Routledge (2016)
William R. Uttal
First published 2017
by Routledge
711 Third Avenue, New York, NY 10017
and by Routledge
2 Park Square, Milton Park, Abingdon, Oxon, OX14 4RN
Routledge is an imprint of the Taylor & Francis Group, an informa business
© 2017 Taylor & Francis
The right of William R. Uttal to be identified as author of this work
has been asserted by him in accordance with sections 77 and 78 of the
Copyright, Designs and Patents Act 1988.
All rights reserved. No part of this book may be reprinted or reproduced
or utilised in any form or by any electronic, mechanical, or other means,
now known or hereafter invented, including photocopying and recording,
or in any information storage or retrieval system, without permission in
writing from the publishers.
Trademark notice: Product or corporate names may be trademarks or
registered trademarks, and are used only for identification and explanation
without intent to infringe.
Library of Congress Cataloging-in-Publication Data
Names: Uttal, William R.
Title: The neuron and the mind : microneuronal theory and
practice in cognitive neuroscience / by William R. Uttal.
Description: New York, NY : Routledge, 2016. | Includes bibliographical
references and index.
Identifiers: LCCN 2015050181 | ISBN 9781138640191 (hb : alk. paper) |
ISBN 9781138640207 (pb : alk. paper) | ISBN 9781315617480
(e : alk. paper)
Subjects: LCSH: Cognitive neuroscience. | Brain—Physiology. |
Neurobiology.
Classification: LCC QP360.5 .U88 2016 | DDC 612.8/233—dc23
LC record available at http://lccn.loc.gov/2015050181
ISBN: 978-1-138-64019-1 (hbk)
ISBN: 978-1-138-64020-7 (pbk)
ISBN: 978-1-315-61748-0 (ebk)
Typeset in Bembo
by Apex CoVantage, LLC
. . . the relationship between brain and cognition is still only poorly
understood. Great progress notwithstanding, neuroscience still cannot
answer the “big questions” about mind and intelligence.
(Olaf Sporns, 2011, p. 179)
To be brutally honest, scientists do not yet have even the remotest idea
of how visual experiences—or indeed any other kind of experiences—
arise from physical events in the brain.
(Stephen E. Palmer, 1999, p. 618)
. . . the initial hope of associating each brain area with a particular func-
tion (Posner, Petersen, Fox, and Raichle, 1998) has not been realized.
(Fox and Friston, 2012, p. 408)
This page intentionally left blank
FOR MITCHAN
This page intentionally left blank
CONTENTS
Preface xi
Acknowledgments xv
1 Introduction 1
Bibliography 183
Index 199
This page intentionally left blank
PREFACE
The brain is arguably the most complex entity that has ever existed. It is made
up of a vast number of cellular components (neurons) whose abundance compares
with cosmological numbers. Recent estimates for the number of cells in the
normal human brain (based on measurements by Ananthanarayanan, Esser, Simon,
and Modha, 2009; Azevedo et al., 2009) are that there are about 86 to 100 bil-
lion neurons in the brain (along with possibly an equal number of glial cells).
Astronomers, for their part, estimate that there are between 10 and 100 billion
stars in a typical galaxy like ours. Thus, the numbers in a galaxy and a brain
are of roughly the same order of magnitude.
Such a comparison is misleading, however. Although the components of brains
and galaxies are both heavily interconnected, the natures of the respective inter-
active forces among their respective components are very different. Stars interact
by means of a single, uniform, and well-described force—gravity—that declines
uniformly with distance. Quite to the contrary, the brain’s neurons are both
locally and remotely interconnected with one another by what other estimates
suggest may be thousands or even tens of thousands of synapses with idiosyncratic
effects. Each of these synapses (or combinations of them) may produce distinc-
tively different interactive effects on a neuron quite unlike the uniform influence
of gravity between stars. Interconnections between the brain’s neurons, therefore,
are not simple or uniform nor are they likely to be as simple as the forces
operating among a galaxy’s stars. Thus, the combinatorial complexity of the
brain’s neuronal network is probably far greater than that of a galaxy!
Given this level of complexity as well as the microscopic nature of neurons
and synapses, formidable obstacles exist to learning about the mechanisms by
means of which tangible brain activity becomes or produces or is intangible thought.
Nevertheless, in the past few decades there have been some extraordinary
xii Preface
technical developments that allow us to begin to examine the brain in ways that
only a few years ago would have been considered impossible. Many of these
developments are just becoming known to both the general public and the
scientific community.
Twenty-five years ago, I set out the goals of my future research program in
an article entitled “On Some Two-Way Barriers between Models and Mecha-
nisms” (Uttal, 1990). In retrospect, I would have titled this article somewhat
differently. I now read much of what I was saying there as a critical appraisal of
neuroreductionism—in particular, how various aspects of neurophysiological
science produce or are related to cognition. This has been the central theme of
my work and is extensively discussed in my recent and not so recent books
(Uttal, 2001, 2005, 2009, 2011, 2013).
I argue here that none of these neuroreductionist theories proposed to
explain cognition at the neuronal level has yet matured to the point that we
can experimentally examine most of their postulates and implications. My
primary goal in this book is to evaluate the current state of theory in cogni-
tive neuroscience specifically at the level of the cellular (i.e., the neuronal)
components of the nervous system. Unfortunately, there are some profound
barriers to progress in this field. What has arisen in the place of solid data-
driven biological-based theory, I believe, are largely metaphors or analogs driven
by available instrumentation—an engineering field where astonishing progress
has been made. An emerging generalization is that our theories are, too a
much greater degree than we appreciate, creatures of whatever technology is
available at any stage of history.
Therefore, it is a secondary goal of this book to survey the new technologies
used by neuroscientists, to explain how they work, and to provide at least an
initial conceptual foundation for the basic principles they highlight. I also want
to consider what parts of the mind–brain conundrum they might help to solve
and which may be completely intractable. Novel applications of these new
techniques and the promise of even more exciting developments in the near
future offer promise of major breakthroughs in our understanding of at least the
anatomical and physiological aspects of brain, if not the “Holy Grail” of modern
brain research: the answer to the question—how does brain activity account for
mental activity?
Make no mistake, by the very act of applying some of these instruments, we
come up against some of the most perplexing problems in modern science.
Some of these problems are practical ones of dealing with the vast numbers of
neurons that are involved. However, others are matters of deep principle that
should be faced before any commitment can be responsibly made to such a
grand enterprise as the proposed billion-dollar investment in the “Brain Activity
Map” (Alivisatos et al., 2013), an approach whose goals are already coming into
conflict with mathematical and physical laws that the brain must obey (Under-
wood, 2014).
Preface xiii
1. Nothing I say here about the difficulties arising in applying powerful new
tools should be interpreted to mean that I am suggesting that we can deter-
mine at this point whether the mind–brain problem is either solvable or
intractable. It is essentially a combined empirical and theoretical question that
may not be answered fully for many years. Of one thing we can be sure—we
are nowhere near even a tentative solution.
2. Should it turn out that the mind–brain problem is formally intractable, such a
finding would necessarily have no special significance outside of the domain
of the natural sciences. That we cannot understand something for practical
reasons cannot be used to argue that there is anything other than materialist
laws at work in the brain as well as in the cosmos. Nor does the absence of a
complete explanation of a complex system such as the mind–brain mean that
we are not able to manipulate that system in productive, practical ways.
3. Furthermore, no one can deny the possibility that some future discovery or
some scientific genius will lead to an overarching theory of mind–brain rela-
tions and overcome what are apparently insurmountable current barriers to
understanding. This book is only a contemporary evaluation of theory and
technology underlying a particular kind of neuroreductionism, the kind we
practice at the beginning of the 21st century. It makes no claims to the future
beyond reasonable extrapolations from existing knowledge and identification
of potential or real obstacles to progress.
4. The fact that there may be formidable technical and conceptual obstacles
lying ahead should not be interpreted as a call for cessation of research at
any level. We are learning much about single neurons and their properties in
complex networks, and, at the very least, some network models or computer
programs have led to useful simulations of human behavior of great value.
The engineering of new tools that simulate human behavior will move on
and whatever limits characterize current technology are likely to eventually, if
not immediately, correct themselves.
5. Nevertheless, there are some fundamental barriers to understanding
how the brain produces the mind that would be well to appreciate. To
attempt to overcome some otherwise well-established physical law, such
as the second law of thermodynamics, or the combinatorial explosion in
computation in an effort to build a perpetual motion machine would be
an enormous waste. We should not ignore the possibility that we might
encounter similar physical or computational limits in cognitive neurosci-
ence and beware of equally wasteful efforts to go beyond what we already
know to be impossible.
xiv Preface
In summary, there are two major questions that I seek to answer in this book.
The first is—are overarching microneuronal theories of cognition possible? The
second is—how does the available instrumentation constrain and dictate what
kind of theories we develop? Should I make some slight progress in answering
these two questions, I would consider my efforts worthwhile.
ACKNOWLEDGMENTS
1.1 Explanation
Of all of the scientific mysteries confronting our inquisitive species, none is more
profound or challenging than understanding how the tangible brain can give
rise to intangible thought. This has traditionally been known as the mind-body
problem but, as we have become more sophisticated, referring to it as the mind–
brain problem seems more appropriate. Despite the enormous amount of research
being conducted that deals with the brain and the mind, it is almost universally
agreed that we actually have made little progress toward resolving the overarching
nature of this relationship. In place of creating anything approximating the grand
theory of cosmological physics, we have learned an enormous amount about
each domain separately but, unfortunately, piecemeal. Fortunately, knowledge
about the anatomy and physiology of the brain is succumbing to powerful new
research techniques, and psychologists have learned much about the ebb and
flow of mental activity by observing behavioral responses.
What we have not done is to have made any substantial progress toward build-
ing a conceptual bridge between the psychological and neurophysiological
domains. The reasons that we have not are multiple and complex; as a result
investigators from many fields have debated whether the problem is in principle
“solvable” or “tractable.” On the one side are the optimists. To them, to “explain”
how the brain makes the mind is just another scientific task, admittedly quite
complicated, but still achievable within the rubric of conventional scientific inquiry.
However difficult the problems may be, the obstacles to understanding them are
just practical matters awaiting the development of a new method or technology.
Most cognitive neuroscientists implicitly accept the ultimate achievability of the
task and assume that however slow it may be in coming, eventually we will
2 Introduction
understand or explain how the brain makes the mind. To think otherwise would
raise serious questions about the raison d’etre of cognitive neuroscience itself.
On the other side of the debate are the pessimists (or, if you prefer, the real-
ists); their countervailing opinion argues that the problem is not just difficult in
practice but is impossible in principle. That is, the complexity of the problem
is such that there is no strategy or tool that we could adopt or invent that would
ever lead to an analysis that would “explain” how the process of converting
brain activity to mental activity takes place. Philosophers, combinatorial mathema-
ticians, some behaviorally oriented psychologists, and even a few of those who
would happily accept the title Cognitive Neuroscientist are now beginning to
accept, implicitly if not explicitly, at least the possibility of impossibility.
The resolution of this dispute, of course, will ultimately come from empirical
studies. Someday we may solve the problem and that will be that—debate over.
Nothing proves the existence of a dimly seen phantom like Bigfoot better than
robust evidence of its existence although absence of such evidence does not
prove it does not exist. In this case, the “existence proof ” for an overarching
neuroreductionist theory would be the demonstration of our ability to construct
high-level behavior or intelligence or sentience or cognition (or whatever it is
that we may wish to call mind) from the properties of low-level neurophysiologi-
cal components. Of course, even then the uncertainty of what amounts to an
acceptable explanation would remain controversial.
This debate sets the stage for this book. In it, I strive to review and evaluate
the current state of a particular kind of mind–brain theory—the kind that is
based on the microscopic, cellular components (i.e., neurons) of the nervous
system. These theories are designated as microneuronal as distinguished from those
that are based on larger chunks of the brain—those designated as macroneural. I
have already dealt with macroneural theories in a previous book (Uttal, 2016),
to which this book is a companion, and will only deal with them casually in
passing in this chapter.
To start, it would be useful to define exactly what we mean by a theory. For
a number of reasons, however, categorizing existing theories is much more arbi-
trary than it may at first seem. Theories come in many kinds, each of which
may emphasize different foundation axioms and postulates. A rough taxonomy
of the major kinds of cognitive and cognitive neuroscience theories would include
the following:
All theories represent hopes for the future; however, each type of theory is
constrained in what it can ultimately accomplish. Not only are there differences
in the level of anatomical complexity at which a theory must operate but also
there are major differences in what relationships should be measured in an
experiment. Thus, the exact theoretical relation between a cognitive process and
a neural response is not always clear even in the most highly correlated data.
Theories also differ in their methodological origins. A “theory” may be built
on any one of the following relations:
The range of theoretical issues with which I have dealt in my previous work
on neuroreductionism can be further clarified by the following list of assump-
tions, properties, postulates, restrictions, and if you wish, biases that have guided
my analysis:
thinking was based on two ideas. The first was that the nervous system’s covert
activity (as evidenced by the resulting overt behavioral responses) could be elicited
by electrical stimuli—an idea famously known since the work of Luigi Galvani
(1737–1798), in which a frog’s leg was activated by an electrical shock generated
by a primitive battery, the Voltaic pile. In addition to suggesting that there was
an electrical basis of nervous action, the efficacy of electrical stimulation meant
that very well-controlled stimuli could be used to generate well-defined responses
in the nervous system.
The second, but far more difficult task, was recording the minute signs of
nervous activity by measuring the electrical signals emitted by organic tissues
and cells. This meant that the entire panoply of electrical measuring instruments
developed over the centuries could be used to study the nervous system at several
different levels of magnification. Although we now know that this electrical
activity can be even more fundamentally explained in the language of ionic
chemistry, specifically of the concentrations and flow of sodium, potassium, and
chloride ions across the neuron’s cell membrane, it was the doorway opened by
electrical recording by the invention of some ingenious devices as the galva-
nometer and oscilloscope that made it possible to understand the operation of
individual neurons. It is still not possible to unravel the complex activity of
networks of many neurons, but progress looms on the horizon as new techno-
logical advances such as parallel-processing computers flood from engineering
laboratories.
The point is that, to an immeasurable but very large degree, both experimental
and theoretical progress in cognitive neuroscience, like all other sciences, has
been driven by available instrumentation. That is, the instrumentation available
at any moment in history largely determines our theoretical extrapolations and
the inferences we draw from our observations. As a result, it is all too easy for
us to be misled by tool availability to overvalue a particular line of theoretical
thought to the exclusion of other more plausible, but empirically unavailable,
alternatives. These biases also extend down into the empirical results themselves
as instrument-based theory also feeds back to guide and direct our experimental
designs and to select to which findings we will pay attention. Thus, for example,
it is argued that many theories may be heavily (if not more so) influenced by
the properties of whatever tool is available than some more germane but less
obvious attributes of the phenomenon under study.
Consider the following example: The most popular of the current tools—the
EEG and fMRI—used by cognitive neuroscientists mainly observe at the macro-
neural level. They record the activity of literally billions of neurons in the
form of cumulative activity localized at particular places (on or in the brain) or
occurring at particular times. This led to a number of prototheories of brain
organization that were couched almost exclusively in the terminology of mac-
roneural brain locations (using fMRI techniques) or temporal sequences (using
EEG techniques) rather than the properties of microscopic networks. This has
10 Introduction
Levels of Analysis
A major barrier to theory building concerns the level at which a problem is
being studied. By level of analysis, I refer to both the scale at which brain activ-
ity becomes mental activity and the scale of our analyses and measurements. In
the present context, a microneuronal level of analysis entails investigating the
responses of cellular and subcellular units such as neurons and synapses and their
respective interactions. A macroneural level of analysis deals with signals that are
inseparable pools or accumulations of these lower-level responses such as fMRI
images or EEG recordings.
Theory development at a microscopic level is severely inhibited for several
very practical reasons. Primary among these, we are overwhelmed by so much
data that empirical-based theory is, for practical reasons, computationally unavail-
able to us. We have no way of both manipulating all of the necessary experi-
mental details and measuring the vast amount of activity in the neuronal network
whose collective but idiosyncratic states must in some sense correspond to cogni-
tive processes. Macroneural approaches to theory are also deeply flawed. They
trade the ability to measure for the convenience of pooled or mixed data, thus
losing the very information—the detailed microstructure of the brain—that they
would need to develop a microneuronal theory.
Thus, at both the microneuronal and macroneural levels, the possibility of
developing an authentic, valid, overarching theory that truly speaks to the question
of how the brain encodes mental activity is currently remote and is likely to be
for the foreseeable future. For these reasons, I suggest that both levels of analysis
are not promising foundations on which to build a theory of mind–brain relations.
However, unlike the microneuronal level, macroneural experiments cannot only
be conceived but actually implemented. This simple fact introduces a strong bias
into the entire conversation. What follows is a more detailed discussion of some
of the barriers that confront the cognitive neuroscience theoretician.
Complexity
It is a truism that the brain is an extraordinarily complex organ. Indeed, given
the idiosyncratic nature of its interconnections, an argument can be made that
it is the most complex entity in the universe. Regardless of which level of analysis
one is considering, the combinatorial complexity of the brain’s neuronal network
stresses any possible explanation of its function to its limits for a relatively simple
practical reason—there are too many neurons and too many interconnections
involved in cognitive processes to be processed by any conceivable information
processing machine. It surprises many how very few interacting neurons it takes
to pose an intractable combinatoric problem. The complexity issue has been
known for years. As long ago as the 1960s, Leon Harmon was demonstrating a
simple three-neuron simulator that produced unpredictable results.
12 Introduction
Inaccessibility
Why should this imprecision of definition and measurement of cognitive processes
exist? A major conceptual reason is that hypothetical cognitive constructs are not
only inherently difficult to define but also impossible to directly measure. From
14 Introduction
any point of view, mental experiences are private, intrapersonal, and inaccessible
psychological states that can only be experienced by the individual and are not
directly sharable with others. Neither introspection nor experimental assays can
provide us with direct evidence of the values, properties, and dimensions of mental
phenomena. Our reasoning concerning “other minds” is, therefore, limited to
analogies based on the Cartesian assumption that “Cogito ergo sum” and the
analogous conviction that our fellows must be just like us. This means that we
must use behavioral surrogates such as “percent correct” and “reaction time” as
indirect measures of cognitive response, thus isolating us from the phenomena of
primary interest to psychologists—our inner private cognitive responses.
Another reason that cognitive variables may be inaccessible is that they prob-
ably reflect the outcome of extremely complex multi-variate neural responses
that cannot be characterized by single (or a few) measures typical of modern
cognitive neuroscientific research. The net effect of inadequate definitions and
inaccessible, underdetermined responses is to place conceptual barriers between
the various levels of representation of cognitive activity and, thus, further limit
the possibilities of developing a theory of mind–brain relationships.
Unfortunately, general disappointment was the main result when the same
protocols and concepts were used and as false analogies were drawn between
sensory and motor information transmission processes on the one hand, and cogni-
tive activities on the other. These optimistic (and probably false) analogies persist
to this day. In short, determining the neurophysiological basis of cognition is a
much more challenging process than is the analogous task for sensory and motor
processes. Progress in one does not necessarily portend progress in the other.
Many theorists point out that they are not talking about the nervous system
as an actual structure undergoing physiological or bio-chemical changes
but only as a system with a certain dynamic output.
(p. 194)
What it appears that many investigators (e.g., Anderson, 2010; Johnson, 2011)
are doing is to develop plausible, but highly speculative, “hypothetical constructs”
that could, in principle, describe the behavior but which do not enjoy the sup-
port of empirical linkages between the psychological and neural mechanisms.
The hypothetical mechanisms are then defined in neurophysiological terminology.
At best, theories of this type are heuristics that are not likely to rise from pro-
totheoretical hypotheses to robust neural theories. Perhaps we should be reevalu-
ating the important insight about the Conceptual Nervous System offered by
Skinner and distinguish between robust neurophysiological entities and “nick-
names” for hypothetical constructs.
two goals, it should be emphasized, are not the same and, for the moment, no
bridge exists between the anatomical and the cognitive.
Each of these two approaches is challenged by technical and conceptual dif-
ficulties. Difficulties with the microneuronal approach (in which the integrity
of the individual neuronal response is preserved) are counterbalanced by difficul-
ties with the macroneural approach (in which the responses represent pooled or
cumulated information that represents cognitive processes). Technologically, these
two perspectives (the microneuronal and the macroneural) are very different in
terms of their basic assumptions and the ease with which their respective dif-
ficulties can be overcome. Macroneural measures are cumbersome, noise ridden,
and often quite arbitrary in their evaluation. They may require elaborate statistical
analysis to extract localized regions of activity that might correlate with cognitive
processes. This is the basic presumption forced on macroneural theorists by the
instruments used to measure them. Although often expensive and time consum-
ing, macroneural experiments are relatively easy to design to test their assumptions
using such devices as the EEG and the fMRI. The main problem with these
macroneural neural measures, however, is that they are cumulative responses
pooling or summating the responses of individual neurons and their intercon-
nections. Thus, much of the detailed information of microneuronal organization
is lost. I have discussed the weaknesses of the macroneural approach much more
completely elsewhere (Uttal, 2001, 2011, 2013, 2016).
Should the competing hypothesis—that the essential aspects of cognitive
encoding are to be found at the microneuronal network level—be correct, then
macroneural representation would be a deeply flawed approach to understanding
mind–brain relationships. Rather than measuring macroscopic blood-oxygen levels
with fMRIs as an indirect indicator of brain activity or summations of electrical
activity from scalp electrodes with EEGs, we should be studying the microscopic
action and interaction of neurons—and in particular the networks into which
they are organized.
Empirical studies of microneuronal events are, on the other hand, much more
difficult to implement experimentally. Most inhibiting to the microneuronal
approach is that we have no specific knowledge of how many neurons are needed
for a network to be “cognitively significant.” It is likely that the number is very
large—possibly as large as the number of neurons in the entire brain—but this
is pure conjecture.
the visual cortex that were selectively responding to visual stimuli.3) From this
point in the history of this powerful tool, the influence of the fMRI technology
on cognitive neuroscience cannot be overstated. Theoretically correct or not,
fMRI technology changed the face of the science and created a zeitgeist in which
the macroscopic measurements of brain activity were to dominate both experi-
mental protocols and theoretical positions for years. Once again, the historical
role of brain-imaging systems illustrates the strong impact that whatever technol-
ogy is available to science can have on conceptual basis as well as its empirical
findings. It was only decades later that some of the conceptual and technical
deficiencies of the macroneural approach began to emerge.
data? Although there has been some progress in selecting alternate sensory
responses from among a limited set of brain images, much of this modest
achievement seems to be attributable to topologically preserved peripheral
sensory encoding. For example, visual stimuli are represented by retinotopic
maps in the primary sensory area that preserve the topology of stimuli and,
thus, may maintain accessible and useful information about the spatial pattern
of a stimulus. This is not possible with the symbolic brain representations
driven by higher-order cognitive processes that have no isomorphic relations
between brain activity and those cognitive processes.
7. Many cognitive neuroscientists, heavily influenced by earlier psychological tax-
onomies, are trying to use brain-imaging devices to determine something about
the localized brain mechanisms by which cognitive processes are carried out.
For example, how do we learn? What brain changes occur in learning? A major
unsolved problem is where and what is the engram? Others seek to understand
the emergence of consciousness, attention, and other vaguely defined high-level
cognitive processes by determining which brain regions are activated when
these cognitive processes are manipulated. This is the main theme of much of
brain-imaging research these days. However, there are compelling reasons to
believe that this simplistic concept of seeking correspondences between cogni-
tive processes and specific brain locations may be ill chosen.
8. Many technical issues occupy the time and energies of cognitive scientists,
not all of which are aimed at the great question of how the mind emerges
from brain processes. There is a continued effort to develop techniques to
extract the best possible and largest amount of data from noisy brain images.
Still other investigators are concerned with the technical matter of how we
can pool or combine methods with low statistical power to produce higher-
power experiments in order to yield more significant data—the meta-analysis
approach. There remains an unresolved question when we use this method—
can we plausibly combine experiments that are often varied in method and
conceptualization? Depending on the actual signal-to-noise relations and
whether a signal actually exists or is a manifestation of a stochastic system,
this may also be an ill-chosen expenditure of resources.
9. Most generally, cognitive neuroscience is currently aimed at establishing the
macroneural neural basis of cognition. In short, the main goal of this science
is to provide some insights into the great question—how does the brain make
the mind? To this overarching question, there is little in the way of either an
answer or a satisfactory theory yet available. It is highly problematic whether
we have even begun to answer this question given the likelihood that macro-
neural techniques such as the fMRI may actually obscure the critical micro-
neuronal information that would answer this question.
Obviously, whereas some of these questions and goals represent issues that are
of existential importance, others are of merely epistemic significance. Equally
Introduction 23
obviously, most of the cognitive neuroscience questions posed here are currently
answerable. For those that can be answered, it is my hope that the subsequent
discussions in this book will at least help to clarify the issues arising.
Like all other empirical approaches directed at a solution of the mind–brain
problem, there are barriers and difficulties that impede progress toward a solution.
Some are general to all cognitive neuroscience strategies while others are specific
to the macroneural approach. One that falls into the latter class is the challenge
faced by the pooled nature of the responses, that is, by the fact that any macro-
scopic measure is a composite of the individual responses of many individual
neuronal responses. Currently, there is a prevailing opinion that it is the ensemble
coding or states of the individual responses that in some ultimate sense must be
understood if we are ever to make progress in unraveling the brain’s role in
cognition. Unfortunately for this approach, the pooled signals (the physical sum
of many neuronal responses) actually have lost the critical information.
In such a situation, in which an exhaustive microneuronal analysis is beyond
our technical capabilities, science tends to turn to available macroneural cumula-
tive measures (e.g., the fMRI or the EEG) and effectively let nature do the
accumulative analysis for approximate solutions to complex neuronal network
problems. The expectation is that the critical microneuronal information will
be at least partially preserved at the macroneural level. How the cumulative
process works neurophysiologically is not completely understood, but it may
involve summations of either local (e.g., Logothetis et al., 2001) or spike action
potentials (e.g., Mukamel et al., 2005).
Currently, it seems likely that the salient microneuronal information is not,
in fact, preserved in macroneural level measurements and, thus, the details of
neuronal responses and interactions cannot be retrieved from cumulative measures
such as fMRIs. The voltages and ionic currents that are the responses of individual
neurons must follow the laws of physics as they are added and subtracted from
each other. Once added or accumulated into pooled responses, basic thermody-
namic principles argue that the original initial responses of a given neuron cannot
be retrieved any more than an egg can be unscrambled. This constraint applies
directly to the pooling of neuronal or hemodynamic responses directly or indi-
rectly into signals like the fMRI or the EEG. Thus, there is an innate ambiguity
about the origins of any macroneural signal based on accumulations of micro-
neuronal activity—a huge variety of different configurations of microneuronal
neuronal activities may produce exactly the same macroneural response. This
means that scores based on differences between macroneural responses—for,
example, the subtractive method—produce data that cannot discriminate between
different microneuronal brain states. This issue is such an important matter that
it has been continuously discussed for almost two decades (Van Orden and Papp,
1997) and is still in need of defense (Roskies, 2010).
The implication of the fact that the same macroneural responses can be
produced by a variety of different microneuronal responses is that macroneural
24 Introduction
fMRI images are neutral with regard to their microneuronal origins. This is the
fundamental weakness of the macroneural approach; the critical microneuronal
information essential to understanding the transition from brain state to cognitive
activity may have been irretrievably lost as a result of the pooling process.
The one possible exception to this neutrality of fMRI images is that informa-
tion may be preserved concerning the respective locations of function-specific
nodes, regions, activation sites, or locales on and in the brain—if they actually
exist. Regardless of the truth or falsity of this fundamental postulate, finding the
brain loci associated with cognitive processes has been the core of the larger
portion of the research carried out with current fMRI equipment. Should it turn
out that the macroneural representations of specific cognitive activities are actually
valid and cognitive processes are localizable to particular nodes or activation sites
on the brain, the location of the respective cognitive responses might be considered
to be preserved information. However, even this most basic postulate of brain
organization—localization—may not be empirically supportable.
Another closely related barrier to macroneural theory building is generically
referred to as underdetermination. Underdetermination implies that there is
insufficient information in the results available from an experiment or group of
experiments to answer the salient questions for which the research was carried
out. Some of this information was lost as a result of pooling as previously
discussed—microneuronal information is not preserved in macroneural signals.
However, it is more likely that most of it is due to the fundamental conceptual
difficulty that the information necessary to answer a cognitive neuroscience
question was never present in available macroneural measurements. Both behavior
and formal mathematical models, for example, are underdetermined in that they
do not contain enough information to uniquely determine the specific underly-
ing mechanisms that produce the behavior (Moore, 1956). Indeed, neither behavior
nor mathematics can distinguish among what may be an innumerable number
of alternative possible and plausible underlying mechanisms.
Furthermore, additional experiments cannot always be carried out that will
resolve this deficiency. Hilgetag, O’Neill, and Young (1996), for example, have
argued on purely mathematical grounds that complex networks such as those
proposed by Van Essen, Anderson, and Felleman (1992) for the primate visual
system cannot be placed in a hierarchy of activation order because the outcomes
are fundamentally underdetermined. Carrying out additional experiments leads,
paradoxically according to Hilgetag and his co-workers, to an increase in the
number of plausible mechanisms at a rate that exceeds the number of new
experiments that can be designed to determine the response hierarchy. In short,
simply collecting more data may not help to resolve issues of underlying mecha-
nism if the data are underdetermined; indeed, it may exacerbate the problem.
Another way to describe underdetermination is closely linked to potential
neural mechanisms. I refer here to the “many to one” constraint. If, for
Introduction 25
• That the brain is organized into regional subdivisions that encode or represent
cognitive functions
• That these regions will exhibit a distinctive local neurophysiological response
when a cognitive stimulus is activated
• That activation of one of these regions by, say, an electrical response will influ-
ence a relevant behavioral, cognitive, or neural response
• That reliable experiments can be reported in accord with robust experimental
and statistical paradigms that permit replication of those experiments
I now review some of the most compelling arguments against these assumptions.
Collectively these experiments suggest that serious questions are raised about
macroneural theories at their most basic conceptual levels.
26 Introduction
Their results indicated that the more data that were included in the analysis,
the greater was the extent of the distributed activity across the brain and less
evidence existed for localized function. To support this result, Thyreau et al.
(2012) reported a similar experimental result in which they also used a very
large subject sample (n = 1,326) and found generally the same result—widespread
distribution of responses across the entire averaged brain for even the most
elementary cognitive process. They pointed out that when very large samples
were used, even small responses began to achieve significance; this expansion
increased until virtually the whole brain was responding significantly. The
important implication of their work, like that of Gonzalez Castillo and his col-
leagues, was that the apparent segregation of the human brain into cognitively
specialized regions associated with particular cognitive processes or brain activa-
tion nodes may possibly have been an artifact of inadequate sample size.
Oosterwijk et al. (2012) also supported the idea that the brain mechanisms
of cognition must be more widely distributed in the brain than is currently held
by localization theorists. From their point of view, the assertion that the neural
mechanisms of cognitive processes are localized in the form of macroneural
function-specific regions scattered across the brain was fundamentally incorrect.
Specifically, they noted:
The point of these studies is that the traditional search for localized macroneural
regions mapping onto behaviorally defined psychological constructs is likely to
be fruitless. The brain is not a system of isolated function-specific nodes; it is
more likely to be a broadly, even universally, distributed system of neural mecha-
nisms that are neither function-specific nor localizable to any particular region.
Such a whole-brain metaphor is vastly different than the current one guiding
much of macroneural thinking about how the brain is organized. In sum, the
general import of this work is that the search for narrowly localized brain regions
associable with cognitive processes may be sterile. The conflict between these
emerging data and the search for localizable neural mechanisms may require a
major change in whatever kind of theoretical model is being pursued.
Although Carp does not assert that all experimental results vary across all
pipelines, what he does argue is that “others varied widely from one pipeline
to another” (p. 12).The fact that some pipelines, even the preponderance of
them, were to some degree in agreement (a point made strongly in Carp, 2012a)
does not mitigate the fact that there was an increased tendency for false positives
to occur as a function of the variability in analytical methodology.
The conclusion to which we are drawn by this cogent critique and the concept
of alternative pipelines is that variation in analytical method might be driving
research conclusions (such as whether they are significant or not) to an unknown
degree. Thus, great uncertainty is injected into even basic decisions about the
significance or non-significance of the raw results. Should our scientific results
vary with method and not converge for different methods, they would not inspire
much confidence in their validity as measures of the underlying neural reality.
Similar points have been made by Simmons, Nelson, and Simonsohn (2011)
and Bakker and Wicherts (2011) for purely psychological studies. Both groups
also noted that there is such enormous flexibility in the way we collect and analyze
data that spurious false positives are encouraged. They emphasized several additional
factors as contributing to poor decisions that lead to error: (1) ambiguity about
the basis for the p < .05 criterion, (2) the arbitrariness of when to terminate data
collection, and (3) simple calculation errors. Bakker and Wicherts, for example,
stated that as many as 33% of the 281 reports they reviewed either misreported
their findings or made an error in calculation.
Introduction 29
1. Intersubject variability
2. Intrasubject variability
3. Absence of sham stimulation and blind analysis techniques
4. Motor and cognitive interference
5. Controls for physical nature of the direct current stimuli
When examined in the light of the equally poor controls and variable results
obtained with the fMRI approach, this critique speaks strongly to the idea that
30 Introduction
As well as:
Although it has long been assumed that emotional, social, and cognitive
phenomena are realized in the operations of separate brain regions or brain
networks, we demonstrate that it is possible to understand the body of
neuroimaging evidence using a framework that relies on domain general,
distributed structure function mapping.
(Barrett and Satpute, 2013, p. 361)
data from experiments will be highly variable even though individual experi-
ments might be significant. Uttal showed that this lack of reliability was per-
meating the entire cognitive science application of fMRI machines by comparing
a variety of experiments. The general effect was to produce variability and
inconsistency despite what must have been acceptable levels of significance.
By 2005, the problem had become widely recognized and resulted in such
hyperbolic statements as “Neuroscience research gets an “F” for reliability” in
a Scientific American blog, “Brain science is drowning in uncertainty” in New
Scientist, and “Brain-imaging is often wrong” in Vox, a “general news site.” More
seriously, submissions of this kind of research to various journals were down,
especially if they dealt with comparisons of cognitive and imaging data. Efforts
to overcome the technical and statistical problems resulted in grants to fund a
“Center for Reproducible Neuroscience” at Stanford University. Although there
may be some corrections to be made in the way data are collected and analyzed,
it is not certain that further investment with what otherwise may be a flawed
approach to the study of brain imaging carried out at the wrong level of analysis
as a pathway to understanding the mind–brain problem is justified.
Over the past twenty years, neuroimaging has been the predominant tech-
nique in behavioral and cognitive neuroscience. The volume of papers and
number of fields it pervades are unrivaled. Despite this, it is curiously
difficult to summarize its achievements in general terms. The simplest
attempts falls [sic] back on two principles that shaped brain mapping at
its inception; namely functional segregation and integration: Neuroimaging
has established functional segregation (the segregated or modular deploy-
ment of functional specialization within brain regions) as a fundament of
brain organization. . . . However the initial hope of associating each brain
area with a particular function (Posner et al., 1998) has not been realized.
While it is true that notions like the “motion sensitive center” and
32 Introduction
“fusiform face area” are part of common imaging parlance, the functionally
informed labeling of all but the smallest portion of cortex remains elusive.
Indeed people now prefer to talk about processing hierarchies, intrinsic
networks and default modes that have no clear association with discrete
cognitive processing components.
(p. 408, italics in original)
Fox and Friston argue that the field has changed or is in the process of
changing and we should no longer attempt to make simple associations between
brain locales and cognitive processes. Although we are likely to continue to use
brain-imaging equipment, they suggest that future work will increasingly be
concerned with the anatomical organization of brain exemplified by the search
for the connectome (e.g., Sporns, Tononi, and Kotter, 2005), the effort to deter-
mine how the brain responds when it is cognitively inactive (e.g., Fox et al.,
2005), or, more generally, “a more physiologically and anatomically informed
approach” (p. 408). Fox and Friston are not alone in drawing this conclusion.
Poldrack (2010), another prominent investigator, concluded that “A review of
the neuroimaging literature suggests that selective association between mental
processes and brain structures is currently impossible to find” (p. 754).
In other words, the field of cognitive neuroscience is gradually changing from
one that sought neural codes for cognitive processes to one more concerned
with the properties of neural networks per se. It seems to be that neuroimaging
is diminishing its activities related to cognition in favor of research that is more
concerned with anatomy and physiology of networks. The reasons for this are
clear:
What a terrible waste of resources resulted from the fact that early critics of
brain imaging as a means of studying the neural basis of cognition were not
attended to (Uttal, 2001; Van Orden and Papp, 1997).
Introduction 33
1. There is increasing reason to support the argument that the mind–brain prob-
lem may be intractable at both the macroneural and microneuronal levels.
Many of these reasons are based on conceptual or logical foundations, but
many others are empirical—simply considering the number of neuronal ele-
ments involved in any mental process raises practical issues of computability
that have no current solutions.
2. Theories come in several different kinds. Not all are neuroreductionist; some
are purely descriptive. However, there is considerable discussion about what
constitutes a neuroreductive theory or whether such an ideal understanding
at any level is actually possible. Scientists still argue over the descriptive versus
the reductive nature of theories.
3. Macroneural theories suffer from the loss of what may be critical data because
of the pooling or summation of microneuronal signals.
4. Microneuronal theories, conversely, suffer from the very large amount of data
that must be processed to answer the mind–brain question. Empirical testing
of the theories may be impossible for the most practical reasons.
5. Obviously, science is driven and constrained by the available measuring
instruments. It is not so obvious, however, that it is not only our empirical
results but also our theories that are so governed. Technology can both enable
and mislead explanations.
Introduction 35
Notes
1. In several places in this chapter I have abstracted material from my previous writing
to review and clarify a necessary point.
2. Some of the material in this section has been adapted and updated from an article by
Uttal (2015).
3. A readable technical discussion of the physics of fMRI recording can be found in Uttal
(2001).
4. This trend may not be continuing; there is some suggestion that the number of papers
that more or less naïvely attempt to localize particular cognitive functions in particular
regions of the brain is slowing. My general impression is that some of our most pres-
tigious journals (e.g., Science) are not publishing articles of the kind that contributed
to this statistic. Indeed, even some of the specialty journals (e.g., NeuroImage) long
associated with correlations of cognitive states and brain images seem to be more heavily
populated with technological issues rather than simplistic assignments of cognitive
processes to localized brain regions than previously has been the case.
This page intentionally left blank
2
SINGLE NEURON PRACTICE
AND THEORY
The idea that a single neuron could also encode more complex concepts such
as a particular face or object persists to this time. However, single neuron theories
remain contentious in the context of today’s network theories—theories that
emphasize the alternative view, namely that the pattern of activity in broadly
distributed, multi-neuron networks is the key to understanding cognitive infor-
mation processing by the brain.
This chapter has two parts. First, it tells the story of those single neuron
research technologies and their strengths and limitations in our search for under-
standing of the mind–brain problem. We start with an elaboration of the methods
and techniques that are used to determine the functions of single neurons. Of
one thing we can be sure—the availability of certain technological instruments
(e.g., high-impedance microelectrodes) had a defining influence on the set of
theories we consider in this chapter. The second part is an evaluation of the
theories that are based on single neurons.
this biochemical and electrical activity. Indeed, hopefully without falling over a
philosophical cliff, we might ask if psychological processes are necessary or just
a meaningless epiphenomenal byproduct of neuronal activity. This may seem
like an extraordinary question; however, considering that a fully functional brain
could in principle function without any consciousness (as a “zombie brain” as
discussed by Chalmers, 1996), the possibility that we are actually automata enjoy-
ing the unanticipated consequence of personal sentience as our nervous systems
go about their business cannot be ignored. The basic problem is, as I have noted,
that we do not yet know what the relationship is between neural activity and
cognitive activity either in terms of single or multiple neurons.
Many neurons specialized for information transmission over long distances have
evolved elongated processes (axons). For example, a single axon runs from the foot
to the spinal cord before synapsing, thus providing a communication line for an
irritation in the toe to be interpreted by the brain as a pain. The axon is an elon-
gation of the neuron’s cell body extending from the cell body or soma, as it oth-
erwise known, to a distant neural locus where it may communicate via a multitude
of synapses to the next neuron in an extensively interconnected network. Other
types of neurons specialized for local information processing may have very short
axons with many connections spread over its surface. In general, axons terminate
in close proximity to the receiving region or dendrite of the next neuron.
Synapses are now known to be of two types, either chemical or electrical.
Chemical synapses conduct by emitting specialized molecules known as “trans-
mitter substances” from the “presynaptic” axonal region of the neuron. These
molecules are received by specialized molecular receptor sites on the “postsyn-
aptic” dendrites. Electrical synapses are rarer, more tightly packed than chemical
synapses, and cannot amplify neuronal signals as do chemical synapses.
Because synapses are so small, their history is much more recent than are
those of other neuronal components. Observation of synapses requires the extreme
magnification provided by electron microscopes—a gain that must be as high
as 2 million compared with the maximum magnification of an optical microscope,
which is limited by the diffraction of light to about 2,000. In the past couple
of decades, ingenious fluorescence techniques (e.g., Hell and Wichmann, 1994)
suggest that new optical technologies may permit “superresolution” beyond what
is imposed by diffraction limits on optical microscopes.
To accelerate the conduction time, elongated axons are usually sheathed in a
multi-layered myelin coating composed of a series of specialized and overlapping
Schwann cells. This myelin sheath is discontinuous with narrow constrictions referred
to as “Nodes of Ranvier” spaced periodically along the length of the axon. The
electrical signals that produce the propagated neural response along an axon are
able to jump from node to node, producing a much higher conduction speed than
would be possible with an unmyelinated fiber. Cells primarily designed for localized
interaction have much shorter axons that are typically not myelinated—indeed,
some presynaptic axons are so short that they are hard to distinguish from the
Single Neuron Practice and Theory 41
Recording Electrodes
The basic role of recording electrodes is to detect and measure the electrical
activities produced by the various components of the nervous system.2 Electrodes
come in a wide variety depending on the specific application. Different electrodes
are necessary for measuring the compound (i.e., pooled) neural responses than
for measuring the isolated responses of single neurons; electrodes that are optimal
42 Single Neuron Practice and Theory
for intraneuronal responses may be quite different than those that are useful for
extraneural ones. Other electrodes are specialized for recording the macroneural
responses such as those picked up by the electroencephalogram (Berger, 1929)
or whole nerve preparations (Dawson and Scott, 1949); these signals may require
only simple electrical contacts—a stainless steel button taped to the skull or
attached to the arm does a satisfactory job of communicating these macroneural
signals to our measuring equipment. In other contexts, there has been a persistent
effort over the years to make direct contact with brain tissue by means of surgi-
cal techniques that use implanted electrodes. (A comprehensive discussion of
electrode technology can be found in an informative article by Cogan, 2008.)
Our main interest in this section is concerned with the electrode types that
can be used to measure single neuron responses. (In Chapter 3, I will deal with
multiple arrays of electrodes that can be used to record from many neurons
simultaneously.)
Neurophysiological recording of the electrical activity of neurons is accom-
plished by measuring the minute voltages produced by shifting ionic distributions
across the cell membrane. To do this, some kind of an electrically sensitive
contact—an electrode—must be placed in or near the neural tissue. Subsequent
amplification of the obtained biological signals to a level that is sufficient to
drive recording equipment is accomplished by appropriate electronic devices.
The amplified signals can then be connected to a display such as a cathode ray
oscilloscope or entered into a computer for subsequent analysis.
The critical and most idiosyncratic component in such a system is the initial
contact point—the actual electrode interface that senses the tiny neuronal electro-
potentials and conducts the information they carry to signal processing equip-
ment. Various forms of electrodes capable of observing single neuron responses
have been used in the history of neuro-electrophysiology depending on the
particular neuroelectrical signal of interest. As we shall shortly see, electrode
technology has driven the course of modern neurophysiology and with it the
theories that have been engendered.
The first attempts to record electrical signals from single neurons were carried
out on surgically excised whole nerves consisting of a bundle of axonal fibers.
The electrodes used to record these signals were typically salt-saturated fiber wicks.
The experimental design was a direct one—whole nerves were surgically dissected
free of their surrounding tissue and laid across the wick; the detected voltage
responses were connected to an amplifier and recorded on whatever apparatus was
available for displaying and recording. Such techniques, however, produced signals
that were the accumulation of the responses of many neurons, a mélange that may
have hidden the activity of the individual neuron, as shown in Figure 3.2.
A step toward single fiber recordings occurred when investigators were able
to dissect single axonal fibers free from a compound nerve. Wick electrodes
were also at first used to detect the electrical activity generated by these single
dissected axons. However, the extracellular responses so recorded were very small
Single Neuron Practice and Theory 43
(a few microvolts) compared with the actual transmembrane resting voltage dif-
ference that was ultimately measured across the membrane; these transmembrane
voltages could be as large as 70 or 80 millivolts or even larger in some cases.
The next important step in measuring neuronal activity occurred when Young
(1936) discovered that the squid Loligo vulgaris had evolved a neuron with a very
large axon—one that could be as large as one millimeter in diameter. This was
large enough to permit a reasonably sized glass tube electrode filled with a
conducting electrolyte solution to be inserted into the cut end of this giant axon
and, thus, into its internal environment. The glass tube, therefore, served as an
intracellular electrode. For the first time, using the accidental availability of the
anatomical anomaly in this model preparation, it became possible to observe the
transmembrane voltage in its full splendor, not just the much diminished version
obtained with extracellular electrodes. When a reference electrode (all voltages
are measured as potential differences between two points) was placed in the
extracellular fluid, the full transmembrane voltage could be measured as the
voltage between the inside and the outside of the neuron.
The enabling “trick” in this case was getting one relatively large electrode
into the interior of an unusually large neuron. Thus, the accident of an evolu-
tionary freak—an unusually large axon—made it possible to develop the ionic
theory of neuronal action. Using the giant neuron of the invertebrate squid L.
vulgaris as a model of vertebrate neuronal ionic mechanisms, Hodgkin and his
colleagues (Hodgkin and Katz, 1949; Hodgkin and Huxley, 1952; Hodgkin,
Huxley, and Katz, 1952) were able to work out many of the details of the ionic
flow across the membrane that accounted for its ability to both create local
potentials and to rapidly propagate a signal from one end of the axon to the
other—research that earned them the Nobel Prize in 1963. It also permitted
them to develop a comprehensive theory relating membrane potentials and ionic
concentrations that persists virtually unchanged up to the present time.
Stimulating Electrodes
Electrodes have to be more than just one-way devices. In addition to their
important uses as sensitive transducers of neural signals, they can, as noted previ-
ously, also be used to stimulate or activate neural tissue. For macroneural stimula-
tion, such as is necessary to evaluate percutaneously the properties of peripheral
nerves (e.g., Uttal, 1959), a simple, hemispherical, stainless steel button, buttressed
by some electrical gel, is all that is required.
Similarly, when stimulation of the brain with large implanted electrodes was
required, the material from which the electrodes were made was relatively
unimportant—almost any nontoxic and corrosion-resistant piece of wire would
be satisfactory. In some situations (e.g., Uttal and Krissoff, 1966), in which a
diffuse spatial stimulus was required, it was sufficient to simply insert two fingers
into separate test tubes filled with saline solution.
44 Single Neuron Practice and Theory
other, un-demagnified end of the glass tube (from which the electrode was pulled)
remained large enough for a conducting metal wire to be inserted to serve as a
reference electrode. The functional role of this intracellular microelectrode is to
convey the relatively large transmembrane potentials to a preamplifier; from there,
signals could be conveyed to display and computer analysis devices.
An early application of the new glass pipette–microelectrode technology to
neurons from the frog’s sciatic nerve was reported shortly after Ling and Gerard’s
publication. The first application of a microelectrode inserted into the human
brain was reported by Ward and Thomas (1955).
An alternative method of making intracellular micron-size microelectrodes is
to electrolytically polish a metal (usually tungsten, platinum, or iridium since steel
points are too brittle), wire down to a 1-μm point, and then coat the wire with
some kind of insulating varnish (Grundfest, Sengstaken, Oettinger, and Gurry, 1950;
Hubel, 1957). The tip is left uncoated so that electrical contact can be made with
the neuron, usually extracellularly. Both forms of electrode—glass micropipettes
and sharpened, then insulated, metal points—thus consist of a tiny conducting
core surrounded by insulation (glass or varnish) that can be used as a membrane-
penetrating, active electrode. The second point of electrical contact, the reference
electrode, can be placed virtually anywhere in the extracellular fluid environment
of the preparation and is decidedly low-tech—a clip or plug can be attached to
any available tissue in the specimen whose nervous system is being studied.
The glass microelectrode is particularly useful for intracellular recording. It is
capable of sensing a maximum amplitude transmembrane voltage from a single
neuron. Insulated electrodes made from many different kinds of metals seem to be
better adapted to extracellular recording, typically resulting in a mixture of neuronal
signals from several nearby neurons. How many neurons will be sensed is determined
by fortuitous accidents of the arrangement and propinquity of the neurons and the
electrode and is typically not under the control of the experimenter.
In sum, microelectrode design is still very much an intuitive art form, each
investigator often constructing their own design to meet the special needs of his
or her experiment. Even the original work reported by Ling and Gerard was
the result of a multi-year effort by several people before their design was stable
and reliable enough to be used in any quantitative manner. For reasons that are
still not entirely clear, minor details in the electrical design or in the material
from which a microelectrode is constructed may have significant effects on the
quality of the signals they detect.
Recording Devices
One of the earliest devices designed to store and display the electrical activity
of the nervous system was Willem Eintoven’s (1860–1927) invention—the
string galvanometer. Although first used to record the electrocardiogram,
eventually it was applied to the measurement of the conduction speeds of
46 Single Neuron Practice and Theory
where Vp is the voltage sensed by the preamplifier and Vm is the voltage produced
by the neuron. Thus, if Rm is much larger than Rp, then Vp (the portion of the
voltage) across the amplifier input will be very small—perhaps undetectably so
Input impedance
of preamplifier
Rp Cp
A
Rm
Input impedance
of microelectrode
Cm
B
Microelectrode
Neuron
by the main amplifier. The net effect of this will be to make the tiny voltage
from the neuron virtually undetectable by the amplifier unless the amplifier input
impedance is itself high. This is the reason that specialized high-input impedance
preamplifiers are so necessary in cellular neurophysiological research.
The second main problem with high-impedance microelectrodes is that they
operate in combination with the capacitance of the system to limit its temporal
response properties in accord with the RC rule:
TC = R × C
where TC is the time constant of the system, R is its resistance, and C is its
capacitance. The TC of such a circuit will determine the overall frequency
response of the recording system; the larger the TC, the less the high-frequency
components of the signal will pass through the system. The required high-
frequency response of a neuron-preamplifier combination must be measured at
least in kHz for neuronal responses to be detected and their waveform reproduced
although some modern experiments may require much higher limits.
Impedance mismatches can also introduce spurious noise of several different
kinds. The engineering of preamplifiers with sufficiently high-impedance prop-
erties to match microelectrodes is an art unto itself, but one that has matured
to such a high level of development that they are currently available off the shelf
from a number of different manufacturers.
There are several basic problems inherent in single neuron theoretical for-
mulation that suggest that it is not the ideal route for theorizing. First, we do
not know how many “concepts” there are in our mental life and so we cannot
tell if even the large number of neurons in the brain is adequate to represent
all of them, each with its own neuron whose activities serve as the psychoneural
equivalent of some mental experience. Counting the number of available neurons
and associating each with its cognitive equivalent, although possible in principle,
probably would not be a practical approach.
Second, unlike the first problem, which could in principle, if not in practice,
be validated by simply finding neurons whose response could be correlated with
a concept, there is no way that we could carry out the necessary experiments
to validate the association between specific cognitions and single neuronal activi-
ties. Although it is reasonable to assume that we could find a neuron that
responded to a particular stimulus, locating a particular neuron whose activation
provided empirical proof of a particular subjective awareness would be, to say
the least, a very challenging task.
Third, practical issues such as the manipulation of a particular neuron (isolat-
ing it, turning it off and on) would always conflate, confuse, and confound the
results of any experiment attempting to show single neuron representations of
cognitive processes.
An important related issue is—how do we account for the selective sensitivi-
ties and responsiveness of individual neurons to very complex “concepts”? In
general, the basic anatomical mechanism purported to explain these findings is
a progressive convergence from the properties of a stimulus onto fewer and fewer
neurons, with the result that more and more complex ideas or concepts can be
represented by fewer and fewer neurons. Ultimately, this convergence approach
would respond selectively to complex gnostic concepts and entities such as “yel-
low Volkswagens” or “grandmothers” or “altruism.” This extends the fundamental
idea of a hierarchy of increasingly complex responses in the visual system origi-
nally suggested by Hubel and Wiesel’s (1968) progression of simple, complex,
and hypercomplex cells. The concept of a convergent hierarchy suggested that
the lower, simpler, spatial-temporal properties defined at the stimulus level (such
as line orientation) might progressively converge with other properties onto
higher-level concepts, eventually as specific as a particular human face, at the
same time that neurons are becoming more and more selective and particular
to fewer and fewer tangible ideas.
The idea of a progressive convergence of constituent codes onto neurons
suggested, to hammer in this point, that the meaning represented by higher-level
neurons becomes more and more specific until complex cognitions such as
particular objects, people, or concepts are represented by the activity of a single
neuron. The important mechanism underlying further development of this idea
is, thus, a progressive convergence of the tuning of a neuron from the spatial-
temporal domain of simple stimuli to the gnostic, conceptual, or symbolic domain.
Single Neuron Practice and Theory 51
At present, however, no one has been able to demonstrate this progressive neu-
rophysiological convergence once one proceeds beyond the peripheral sensory
systems.
In sum, the core of single neuron theory is that once it has been established
that a neuron has a particular sensitivity (such as responding to an individual
face) subsequent activation would be the necessary “cause” or “equivalent” of
the mental experience of that face. The basis of a pure single neuron theory,
therefore, is that the activation of a single neuron is the psychoneural equivalent
of entities of unlimited complexity.
Without a doubt, a statement that asserts that single neuronal activation is
tantamount to experience is replete with uncertainties, ambiguities, and unresolved
subissues. What are the other billion neurons doing while the critical one is
determining what we momentarily perceive or think? Is there some kind of
priority that one neuron has over all others? How did the neuron acquire the
particular sensitivity? Is the individual neuron itself conscious or its response
sufficient to determine a unique conscious experience? Is the role of psychoneural
equivalent being conflated with its equally plausible role as a “trigger”? For
example, if activation of a single neuron is simply acting as a trigger for a more
complex response actually embodied by a network of multiple neurons, then it
might well appear to an experimenter to be causal, but any such attribution
would be fallacious. In short, the single neuron theory, although superficially
simple, is actually extremely subtle and complicated by questions rarely asked
and almost never answered.
The antithesis to the single neuron theory is the neuronal network approach.
Rather than ascribing cognition to the activity of a single neuron, this alternative
proposes that the psychoneural equivalent is embodied in the interactions of a
network of many neurons, all of which are interacting with each other in com-
plex and dynamic manners. Controversies in this case concern how many neurons
are involved in a network and whether we can quantify network activity in a
mathematically tractable form. The raw conceptual simplicity of the single neuron
approach (neuronal response X is the equivalent experience Y) is an advantage
not shared with neuronal network theory. Even at the earliest stages of proto-
theory development describing how brain activity might make the transition
from neural function to psychological experience, preliminary anatomical evi-
dence made it clear that the process would be an overwhelming one that was
not likely to be solved in the foreseeable future if phrased in terms of the network
interactions.
Both single neuron and network theories propose an answer to the question—
what activities and properties of the brain produce, cause, or become cognition?
Both, however, raise many issues of meaning and interpretation that have to be
considered in detail. In subsequent chapters, I consider neural network theories
in detail; here I now discuss the practical and conceptual implications of single
neuron theories. However, before going on, there is one important clarifying
52 Single Neuron Practice and Theory
comment to be made—both the single cell and the neuronal network theories
have a common feature. That is, both maintain information about the status and
activities of individual neurons. Thus, both fall into the microneuronal category.
Although this may not seem to be the case when one observes the massive
neuronal networks that characterize the brain, this is their explanatory level. The
differences with the macroneural approach are profound.
idea that a single neuron could be responsible for representing complex perceptual
experiences became extremely popular, if not dominant, in mid-20th century
thinking.
Evidence provided support for the idea that cortical-level neurons become
selectively responsive to particular human beings (e.g., the well-known, but
hypothetical, “grandmother” neuron originally proposed by Lettvin, as cited
by Gross, 2002). Although some controversy exists about the robustness of
these findings (see the work of Tanaka, 1993, who showed specialization to
face-like geometrical forms rather than particular faces), there does appear to
be a relatively strong body of empirical evidence that suggests that these
observations should be taken very seriously—specifically, the not-to-be-
overlooked idea that there are neurons in the brain that respond only when
very specific and very complex stimulus conditions are met. These highly
specialized neurons, as I suggested earlier, are the end product of a process of
progressive convergence that starts with sensitivity to the physical dimensions
of a stimulus (e.g., its temporal-spatial properties) and progresses to neurons
whose responses are related to the highly particularized meaning of a stimulus
(e.g., a specific object, person, or idea).
A summary of a broad swath of current single neuron theories has been
provided in an important review of the field by Spillmann and Werner (1990).
In it, a distinguished group of investigators described their efforts to link the
neuronal properties of the sensory pathways with perceptual experience—that
is, to assume that these neurophysiological responses are the psychoneural equiva-
lents of cognitive experiences. However, even then, there were voices calling out
for a more thoughtful interpretation of the relation between these peripheral
neuronal codes and perceptual experience. For example, Westheimer (1990)
argued that:
Once the first discoveries were made . . . single cell researchers in the
mammalian visual system developed categorizations which do little for
psychophysicists.
(p. 7)
The problem of how far up the sensory pathway one must go before we
find neural mechanisms that are the equivalent of cognitive experience (rather
than just communication codes) is rarely raised by cognitive neuroscientists.
An exception is the work of Crick and Koch (1995), who asked the question,
“Are we aware of neural activity in the visual cortex?” (p. 121). They sug-
gested that there is no awareness of neuronal activity in visual area V1 but
rather, that V1 is actually better considered as a part of the peripheral trans-
mission system. If their suggestion is correct, then the many conclusions of
cognitive encoding by single neurons based on correlations between psycho-
physical and neurophysiological responses in V1 may have been seriously
misinterpreted. The correlated neuronal responses would actually be reflecting
transmission codes, even as high as V1. Thus, the idea that neuronal responses
in V1 were the psychoneural equivalents of our perceptions would be based
on a false analogy—isomorphism of the dimensions of the stimulus and the
psychological responses. This issue is especially important in the present context
because any support of the cognitive efficacy of a single neuron’s responses in
V1 would become meaningless. Indeed, since the preponderance of the empiri-
cal work supporting single neuron theories has been carried out in the peripheral
systems, most of it does not actually deal with the neuronal coding of aware-
ness or experience per se.
Psychoneural equivalent codes at the level of cognitive activity are the answers
to a different question than the one concerning transmission codes. The equiva-
lence question asks—how do the neurophysiological responses become complex
perceptual phenomena? If Crick and Koch are correct, although there may be
correlations between neuronal activity and cognitive experiences, the relationship
is not a causal one. That is, the neuronal responses are not the psychoneural
equivalents of the experience but are conflations—both being driven by the
common stimulus, but without any direct relationship between the neuronal and
the psychophysical.
Without question, great progress has been made in the sensory domain and,
for all practical purposes, most codes for the key sensory dimensions (i.e., quality,
quantity, spatial extent, and temporal pattern) of most sensory stimuli in a variety
of animals have been broken (Uttal, 1973). If explicating the transmission codes
can be considered to be equivalent to theory development, then this is the most
compelling argument that successful neuroreductionist single neuron theories
have been forthcoming. However, this is not the same thing as demonstrating a
solid link between individual neurons and higher levels of experience. Far less
has been accomplished in understanding the relationships between complex
cognitive processes and their underlying neurophysiological equivalents. Further-
more, if Crick and Koch are correct and transmission mechanisms (as opposed
to cognitive ones) extend much farther up the ascending pathways into the visual
areas of the brain than has been appreciated, then the gap between neuronal and
cognitive responses remains as broad nowadays as it has been for decades.
Single Neuron Practice and Theory 55
. . . consistent with the hypothesis that single neuron activity in the human
amygdala might play a role in either encoding or computing stimulus
values at the time of choice.
(p. 336)
mental experiences. Some of the factors that have engendered this kind of single
neuron theoretical thinking include:
How these neurons arrive at their high degree of stimulus selectivity is pre-
sumed, by analogy to Hubel and Wiesel’s geometrical properties in the periphery,
to be the result of progressive convergence of less selective stimulus properties
from one level of neural representation to more specialized and selective ones at
higher levels. Because it is not likely that specific sensitivities such as a “rabbit,”
“Albert Einstein,” a “yellow Volkswagen,” or any other entity never previously
60 Single Neuron Practice and Theory
the stimulus and the neuronal response would be cognitively irrelevant and
tantamount to committing Yule’s classical error of attributing causation to cor-
relation. The specific and selective neuronal response, in such a case, would
merely be an epiphenomenon of huge amounts of information that converged
on a particular neuron.
The idea of a stimulus-correlated neuronal response without cognitive sig-
nificance is a very different interpretation than one based on Speculative Postulate B—
the idea that activation of such a neuron is tantamount to experiencing its
meaning. In that context, a specific cognitive neuroscience (neuron to cognition)
connection would have been made as opposed to just a neuroscience connection
(stimulus to neuronal activity). To put these questions in other ways—is the
response of a single neuron, which is tuned to a complex concept, the necessary
code for the experience of that concept? That is, is it the psychoneural equiva-
lent? If the answers to such questions are affirmative, we would have come a
long way toward the solution of the mind–brain conundrum. Unfortunately,
since Speculative Postulate B is beset by practical considerations, effective testing
of single neuron theories of cognition seems unlikely in the near future.
This is not the place to expand upon the details of each of these arguments, but
there are some general inferences that can be extracted from Hickok’s analysis
of the frailties of this particular single neuron theory of motor behavior. These
include the following (after Hickok, 2009):
• The primary criticism Hickok raises of this approach is that it is not sup-
ported by empirical evidence. A considerable amount of subsequent research
has failed to find the associations between perception of motor action and
neural responses.
• The hypothesis of “action understanding” neurons is not unique. The idea
of “mirror” neurons is not necessary and thus Ockham’s razor should be
invoked.
• Many of the so-called supportive experiments actively dissociate critical rela-
tionships between neural and behavioral parameters.
• When these neurophysiological findings failed to replicate, the crutch of
mirror “systems” as opposed to single neurons was invented to patch up the
failing theory.
• Despite these problems, the idea of neural representations of action under-
standing persists. (Action understanding is a new version of traditional motor
theories of cognition in which experiences are tied to actual but subliminal
motor responses.) This reflects the age-old problem faced by psychology of
poorly defined attributes or faculties.
1. The conceptual leap from single cells to cognition is too long. It seems illogi-
cal that the vast and intricate neuronal networks should not play a key role in
representing complex ideas.
2. Neural network theories are actually simpler than single neuron theories
given the perplexities and unavoidable complications that follow single neu-
ron theories.
3. Despite the huge number of brain neurons, it seems unlikely that every object
of the external world would have its own neuronal representation.
4. Even if they did, the question of how selectivity for objects and concepts that
did not exist emerges during evolution remains salient.
5. Many cognitive processes do not have simple topological relationships to the
attributes of neural responses and, therefore, simple analogical isomorphism
is not likely to be a satisfactory criterion in determining the psychoneural
equivalent of a concept such as “democracy.” Furthermore, there are some
stimulus dimensions, such as stimulus chromaticity and responses such as per-
ceptual color experience, that cannot be linked together directly by any kind
of straightforward dimensional isomorphism. As a result we must be able to
search for symbolic relationships and interactions—information processing
mechanisms that are executed by the configuration of complex networks
rather than by single neuron states.
6. Perceptual responses are dependent on strong interactions between both the
dimensions of the stimuli and the meaning of the perceptual experiences;
interactions of this kind are not accounted for in single neuron theories.
64 Single Neuron Practice and Theory
7. The sensitivities reflected in the neuronal responses are actually not that good
a match to the concepts being encoded. As noted, what appears to be a face
coding neuron may actually be one responding to primitive shapes (e.g., the
t-shape implicit in the eye-nose-mouth configuration) that are inherent in
face-like stimuli.
8. Artifacts are common—artifacts that are produced by partial properties of the
stimulus.
The “distributed” (i.e., neuronal network) theory (which is based on the idea
that neurons in many parts of the brain are responsible for our cognitive experi-
ences) does not have such a simple expression—from its perspective, the simple
one (neuron) to one (concept) idea is incorrect at its most basic level. Be not
misled, however; network theories confront us with the problem of how to
empirically analyze the actions and interactions of the multitude of neurons that
must be operating in concert to represent a particular concept. This poses a
combinatorial problem of network organization that is very likely to be com-
putationally intractable—but by involving all brain neurons in the process, it
finesses the “exclusivity” problem.
Single neuron theories also generate a false need for some kind of binding
process in which the responses of many neurons must be melded into the
single unified experience. However, when one reevaluates this idea, it seems
obvious that the distributed network hypothesis requires no such binding; all
neurons are incorporated and required to collectively produce the psychologi-
cal response. The state of the network—the configuration of all of the neurons
involved in the network—is itself the psychoneural equivalent. No longer do
we have one neuron equaling one concept but in its place we have the state
of an extended multi-neuron network that offers few opportunities for a
detailed analysis.
As a result, research attention has gradually shifted away from single neurons
encoding both simple and complex properties to networks and macroneural
brain locations as new devices (such as the microelectrode array and the fMRI,
respectively) have offered alternative measures of nervous system activity and its
relation to cognitive processes. Because network theories of various kinds rep-
resent the current consensus, in my opinion, most contemporary cognitive
neuroscientists look upon “grandmother” neurons and their ilk as antique curi-
osities. One reason they have not received more attention is that many of the
critical experiments are not executable in practice and in principle. Should one
be interested in considering single neuron theory further, reviews by Sevush
(2006) and Bowers (2009) present the opposing view. Our attention now shifts
to the antithesis of the single neuron theory—the technology and theory that
have developed to deal with multiple neurons interconnected in networks of
varying degrees of complexity.
Notes
1. This is an excellent example of how the basic chemical properties of a molecule can
lead to the production of organized cells and tissues. Any philosophy that assumes that
there is a neural foundation for mind must assume that such natural processes also
guide organic evolution.
2. Although the electrical activity of a neuron is a useful indicator of its activity, it is
important to note that these electrical signals are really epiphenomenal. The real
mechanisms of neuronal activity are chemical, specifically ionic and molecular, which
66 Single Neuron Practice and Theory
create electrical actions as the chemical activity takes place. It is far easier to measure
electrical activity than it is to measure chemical activity and, thus, much of the discus-
sion of neuronal action is presented in electrical terminology.
3. There is considerable controversy concerning the first use of glass, electrolyte-filled
microelectrodes. Galambos (1996) suggests in a history of neurosciences in his biography
that “glass micropipettes” may have been in use as early as 1910 (p. 187). Thus, their
use was certainly being explored elsewhere around the time Gerard and his group
published their methods.
3
MICRONEURONAL NETWORK
THEORIES—TECHNIQUE AND
METHODS
4096
possible
output
states
1 2 3 4 5 6 7 8 9 10 11 12
12 bit register
IN
FIGURE 3.1 A 12-bit register can represent a much larger number (4096) of output
states.
Microneuronal Network Theories 69
argument—we do not yet know how brain becomes mind. This lack of progress,
fortunately, has not obstructed thinking about neuronal networks. The starting
point for a consideration of a microneuronal theory of how the brain works is
Donald Hebb’s famous conjecture, the next topic in our discussion.
Any frequently repeated, particular stimulation will lead to the slow develop-
ment of a “cell-assembly,” a diffuse structure comprising cells in the cortex
and diencephalon (and also, perhaps in the basal ganglia of the cerebrum),
capable of acting briefly as a closed system, delivering facilitation to other
such systems and usually having a specific motor facilitation. A series of
such events constitutes a “phase sequence”—the thought process.
(p. xix)
1. All aspects of mind, cognition, consciousness, or whatever else you may wish
to call mental activity, are the results of interneuronal processes and mecha-
nisms. These processes and mechanisms are to be found at the microscopic
level of neurons, not larger anatomical units.
2. Specific patterns of neuronal activity are the “psychoneural equivalents” of
specific cognitive processes. Specification of the detailed state of participating
neurons produces a particular cognitive process.
Microneuronal Network Theories 71
1. Identifying each relevant neuron and all of their interconnections that are
involved in a given cognitive process
2. Individually stimulating each of these identified neurons with a coherent
pattern of energies tailored to each neuron’s needs and roles in the relevant
network encoding the cognitive process
3. Recording individual responses from every neuron to determine its individual
and specific response in the relevant network
Obviously, carrying out these tasks would be a technical tour de force of the
first magnitude given that our ability to develop microelectrode arrays suitable
for stimulation or recording has so far been limited to a few hundred (e.g., Wark
et al., 2013). Furthermore, the amount of data obtained by even a few hundred
microelectrodes would tax available storage capabilities.
Should some extraordinary technical development allow us to carry out these
demanding experiments, we would then have to determine how well the criteria
for associating cognitive and neurophysiological measures proposed by Martin,
Grimwood, and Morris (2000) are satisfied. Meeting their four criteria would
authenticate that a microneuronal network response is the causal representation
of a cognitive process; in other words, a theory of how mind emerges from
brain.
Because of these technical and procedural limitations, the brute force, micro-
neuronal strategy of stimulating and recording from a realistically large network of
individual neurons is probably not implementable and can only be conceived of as
a hypothetical Gedanken experiment. Even if by some technological magic we
were able to produce the necessary technical instruments, constraints on data accu-
mulation and analysis would overwhelm any conceivable computer solution to the
mind–brain for simple combinatoric reasons. Thus, there is a glaring inconsistency
between the microneuronal level at which neurons really operate and the macro-
neural level of data obtained at the level of current brain-imaging technology.
However, the second part of Hebb’s conjecture—that it is synaptic efficiency
that controls the state of the network (i.e., producing particular cell assemblies
and, thus, particular cognitive states)—has been tested, in general if not in indi-
vidual neuronal detail. Indeed, this is what has been accomplished by Nabavi
et al. (2014) in an extraordinarily important experiment during which this group
activated and deactivated a brain circuit purported to control a behavioral response
by means of genetic manipulations. Specifically they were able manipulate syn-
aptic activity in the form of LTPs to establish a conditioned fear response and
then reverse this “cognitive” state by producing a Long-Term Depression (LTD—a
reversal of the potentiation) with different optogenetic stimuli. This was accom-
plished by using optogenetic stimuli to alternatively activate and deactivate
synaptic activity, respectively. This procedure permitted Nabavi and his colleagues
to enhance (LTP) or reduce (LTD) activity in the amygdala—a region of the
brain associated with conditioned fear in the rat—by applying two different
trains of patterned light as opposing stimuli.1
The extraordinary outcome of this experiment was that by using this tech-
nique, the rat could be controlled to exhibit a conditioned response analogous
to fear under the LTP conditions, but behavioral passivity under the LTD condi-
tions. Furthermore, and most interesting, was that these researchers were able to
reversibly turn the fear behavior on and off by alternating between the LTP and
LTD synaptic conditions. The proximal independent variable in this experiment,
therefore, is the degree of conductivity of the collective responses of many syn-
apses in a region (the lateral amygdala) of the rat brain thought to be associated
with a well-known behavioral experiment: The dependent variable was a specific
behavior. This experiment provided the best evidence yet obtained to support
what only had been a hypothesis for many years—that the synaptic activity
regulated the behavior, thus supporting the Hebb conjecture.
Although we still do not know the exact synaptic or neuronal states at the
individual neuronal level for this model cognitive process, Nabavi and his col-
leagues have now shown that manipulation, albeit collective, of the degree of
synaptic potentiation produces correlated (and possibly causally related) changes
in an observable behavior. This is solid support for the parts of Hebb’s conjecture
that state that, in general, synaptic processes regulate neuronal states and thus
behavior. Given all of the challenges in this complex field of research, this is a
Microneuronal Network Theories 75
remarkable achievement. This experiment meets the first, third, and fourth criteria
enunciated by Martin, Grimswood, and Morris. The one criterion that it does
not satisfy is mimicry, which requires information about specific patterns of
individual synaptic conductivity—information that still remains elusive.
What Nabavi et al. accomplished, to sum up their conclusion, is that synaptic
states can alter behavior by turning on and off a particular synaptic configura-
tion, thus confirming one of the prime assertions of Hebb’s conjecture. In their
words:
Of course, this does not speak to the other part of Hebb’s conjecture—that is, how
the detailed pattern of the neuronal and synaptic states are decoded into specific
cognitive states. That would require meeting the mimicry criterion, a test that is
not applied in the Nabavi et al. experiment and may not be possible to satisfy.
Why should it be so difficult to provide an answer to the first part of Hebb’s
conjecture and to develop a microneuronal network theory of cognition? The
answer to this question is that some of the very properties that make the Hebb
conjecture plausible and logical also make it intractable to conventional labora-
tory experimentation. These properties include:
see, just increasing numbers sometimes produces its own challenges. The need
for the ability to examine the function of many neurons simultaneously is obvi-
ous given the polyneuronal basis of current neuronal network theory.
However, there are several caveats that have to be made before we embark
on a discussion of the development of MEAs. First, no matter how many elec-
trodes may be combined into an MEA, the number of electrodes (and, thus the
number of neurons from which we may record) will probably always be smaller
than the enormous number of neurons probably involved in any cognitive
process.
Second, as we succeed in developing the technology to record from more
and more neurons, we will be facing an increasing counterforce—the increasing
combinatoric complexity and skyrocketing computational and storage require-
ments to analyze even modest-sized networks.
Third, despite some ingenious use of miniature manufacturing techniques
originally developed for computer chip fabrication, MEAs, in general, consist of
electrodes that are still too large to isolate the ultramicroscopic details of a cog-
nitively significant neuronal network. This is a technological issue that may be
overcome in the future, but, for the moment, represents a serious barrier to
progress.
Fourth, MEAs, which are almost universally extracellular, consist of electrodes
that connect to surrounding neurons in a way that is currently not predictable
or regular, but at least quasi-random. Furthermore, even if the connections were
deterministic rather than stochastic, it is by no means certain that the intricately
interconnected neuronal network of many neurons would map directly from
the MEA to the nervous system. Distant neurons may produce large responses
that are incongruent with the geometrical array implied by the MEA. Alterna-
tively, nearby ones may produce fortuitously small responses for what are usually
obscure reasons where one would have expected large responses. Thus, the goal
of inferring neural network organization from the observed responses may not
be achievable—we may not be able to solve the clustering or sorting problem
(to be introduced shortly) in which responses are associated with specific neurons
or to achieve our goal of determining the connections between particular neu-
rons. The problem of limited detailed knowledge of electrode-neuron connectivity
is exacerbated by the dynamic, transitory nature of neuronal networks.
In general, we do not have adequate control over which neurons will be
connected to which electrodes. Thus, associating the geometry of a particular
neuronal network with cognitive processes remains a challenging task. Therefore,
MEAs are likely to tantalize us by providing data that is conceptually relevant
(cognition is a product of a microneuronal network) but that will never be able
to fully satisfy criteria such as those proposed by Martin, Grimwood, and Morris
(2000) for robust empirical proof. Make no mistake, we are certain to learn
wonderful things about both the nervous systems and cognition, but, for the
moment at least, what seems currently promising may just be another dead end
78 Microneuronal Network Theories
in our quest to explain how brain matter becomes or produces intangible expe-
rience. For these reasons, it is unlikely that even the most advanced MEA will
ever be able to provide a foundation for a detailed explanatory theory at this
microlevel, however promising the idea may be. It may be necessary to concep-
tualize the problem in some completely different way that we cannot currently
anticipate to break the neuronal network–cognitive code.
Although it is currently impossible to simultaneously record from enough of the
enormous number of neurons that must be involved in even the simplest cognitively
relevant network, efforts have been made to construct assemblages of electrodes so
that the activity of a larger number of neurons can be recorded simultaneously than
is possible with a single microelectrode. Currently available MEAs consist of any-
where from a few to several thousand electrodes. The guiding presumption is that
by observing the sequence of activations or correlations among a modest number
of neurons we may achieve some insight into the basic properties of more realisti-
cally sized neuronal networks. The following section details some of the technological
progress that has been made in producing MEAs in the past decade or so.
FIGURE 3.2 An early form of a multiple electrode array (MEA) formed by bundling a
group of wire electrodes (a) and the typical irregular and mixed responses recorded
with such a bundle (b).
Reproduced from Kruger and Bach (1981) with permission.
method for producing high-aspect (in which the length of the electrodes con-
siderably exceeds its width) MEAs so that the electrodes could be pushed deeply
into slabs of mouse hippocampus. This large penetrating depth was required
because the neurons of interest in their experiments were located deep in the
brain. Kibler and his colleagues used arrays consisting of as many as 64 electrodes.
Spira and Hai (2013) present a comprehensive discussion of these new develop-
ments in penetrating multielectrode technology and the problems they raise, as
well as their advantages for the simultaneous recording of multiple intracellular
responses.
However, there remain profound difficulties in using such an array. The simple
act of mechanically connecting large numbers of multiplexed preamplifiers to such
an array is a challenging task unto itself. As we see, most modern research is still
limited to the analysis of relatively few neurons compared with the number of
neurons that must be involved in even the simplest kind of cognitive process.
The main problem with such an MEA, as mentioned earlier, is not
technological—indeed, advances in instrumentation, especially wireless transmis-
sion and microengineering procedures, are likely soon to overcome many or all
of these handicaps. Instead, it is the vast amount of information generated by
even a modest number of electrodes that may define the limits of their application.
What the future holds, of course, cannot be predicted and there has been
continuing progress in increasing the number of available electrodes, if not the
ability to process the large amount of data they could conceivably generate. In
particular, Eversmann et al. (2003) produced an MEA consisting of 16,384
electrodes arranged in a regular array of 128 × 128 individual electrical contacts.
Their MEAs were fabricated from complementary metal oxide semiconductor
(CMOS) technology. Figure 3.5 shows a small portion of Eversmann’s electrode
array and some disassociated neurons that have been deposited on it.
To the best of my knowledge, MEAs of this size have never been used in a
real experiment beyond limited demonstrations of the advanced technology
the exception rather than the rule. As we see later, one of the ultimate goals of
MEA research is to develop analytical techniques to separate mixed neuronal
responses and associate them with individual source neurons.
Shlens et al. wisely concluded that their observation of the relatively simple
pattern of organization in the retina would probably not be generalizable to
more central brain structures. Citing the work of Yoshimura, Dantzker, and
Callaway (2005) on the nonrandom network properties of the multilayer orga-
nization of cortical columns, they suggested that their ability to work out the
detailed network interactions in a “simple” structure such as the retina may not
be possible in more central structures.
MEAs have also been useful in determining the toxicity of various chemicals
on neuronal activity by examining the overall diminishment of neuronal activity
(Steidl et al., 2006). In their experiment, excised slabs of hippocampal tissue were
perfused with antagonists that specifically inhibited synaptic transmitter activity.
MEAs have advantages over single electrodes in such an application because they
can measure system-wide effects in multicellular preparations that may not be obvi-
ous in single cell responses. Furthermore, the specific interconnections of the neurons
from which they recorded were unimportant in light of their goal to determine a
system-wide effect. A comprehensive and up-to-date discussion of the use of MEAs
in neurotoxicity evaluations can be found in Johnstone et al. (2010).
The search for properties of central neuronal networks has been limited so
far because of their complexity. Currently, in most cases, only the overall or
average properties of a network can be determined or the data made available
only in some other cumulative form. Although it is possible with current MEA
instrumentation to observe the activity of a number of neurons simultaneously,
it is far more difficult to identify the sequential dependencies and, thus, inter-
connections of their activities, and, therefore, the specifics of their network
organization.
One very germane example of this kind of research that highlights both the
advantages and disadvantages of MEAs is the determination of the bursting activity
of cultured neuronal tissue. It is in this study that the capabilities of the MEA
technique are currently most clearly demonstrated. It is also in this context that
the great weakness of applying the MEA technique system is also made clear.
Bursting activity of neurons has been observed for many years (e.g., see
Murphy, Blatter, Wier, and Baraban, 1992). It is particularly noticeable in in
vitro clusters of cultured neurons. However, it can also be observed in intact
slabs of neural tissue. Bursting seems to be a ubiquitous property of isolated
neural tissue and may be due to an absence of inhibitory stimuli from other
parts of the brain or, quite to the contrary, as a result of exogenous stimulation
(e.g., see Wagenaar, Madhavan, Pine, and Potter, 2005).
The problem with any measurement of the activity of multiple neurons with
MEAs is that it is difficult to determine the sequential order (or hierarchy of
firing) in which the multiply interconnected neurons respond (a problem high-
lighted by Hilgetag, O’Neill, and Young, 1996). This is especially true in in vitro
preparations in which virtually all of the neurons in a specimen may ultimately
respond simultaneously and spontaneously. This near universal activity negates
88 Microneuronal Network Theories
one of the most promising hopes of those who use MEAs—the expectation that
they will allow us to examine the significant interactions of a network of neu-
rons. As a result, although there have been some associations of bursting activity
with behavioral learning, no obvious relationships have been observed that have
plausibly related bursting activity to controlled stimulation. In fact, in one of
the most notable (for reporting a negative result) articles on this problem using
the components of an experiment we have just been discussing (MEAs, cultured
neurons, and bursting activity) Wagenaar, Pine, and Potter (2006) reported no
difference in the development of bursting activity when stimulated and when
spontaneous. What we are left with in such situations is a dearth of information
describing the detailed properties of neuronal connections. Others (e.g., see
Belykh, de Lange, and Hasler, 2005) argue that synchrony of bursting neurons
depends only on the general topology of the network rather than the specific
interneuronal interactions. Specifically, they suggest that the synchrony of burst-
ing brain tissue depends solely on the number of inputs to a neuron and not
the specific details of the network. This interpretation suggests that “bursting”
may be an abnormal endogenous condition that occurs when some threshold
number of neurons is active. The obvious analogy, of course, is to seizure activity
in intact brains.
The inescapable conclusion one must draw from the extreme instability
exhibited by the bursting phenomenon is that not only is the brain complex,
but also it is not stable over time. Brain networks seem to reorganize themselves
for mysterious reasons in a way that may preclude any notion that we may be
able to decode a consistent neuronal network, much less a stable experimental
environment. Psychologists well understand the variability of our thoughts. Any
experiment, therefore, that seeks to correlate cognitive and neurophysiological
findings may be shooting at a moving target in a way that makes the goal of
understanding the network of neurons that makes up cognition impossible.
A useful and practical recent application of MEAs has been their use as
stimulators in sensory prosthetic devices that serve as supplements to deficient
vision and hearing. Among the most successful applications of this technology
has been in the auditory system in the form of chronic cochlear implants. Among
the first to experiment with this idea in a surgical context were Djourno and
Eyries (1957). In their procedure, electrodes of the kind shown in Figure 3.6
were inserted through the oval window directly into the snail shell–shaped
cochlear space following along its curving course. At various points, an electrode
shorn of its insulation permits electrical stimulation of the remnants of the
cochlear nerve and any surviving hair cells.
Cochlear implants have been remarkably successful, as anyone who has spoken
to a person with such a device will quickly appreciate. Evaluated behaviorally,
the success of such a system is really remarkable and conversation can take place
almost normally. Clinical evaluations by several laboratories generally show 85%
sentence comprehension (see the review by Zeng, Rebscher, Harrison, Sun, and
Microneuronal Network Theories 89
Feng, 2008), a notable improvement over patients who lost hearing due to hair
cell and early acoustic nerve damage.
The high level of success with cochlear implants is largely due to the one-
dimensional and orderly spatial manner in which the various acoustic frequencies
are linearly encoded along the cochlea—a discovery for which Georg von Bekesy
(1899–1972) received the Nobel Prize in 1961.
MEAs such as those shown in Figure 3.7 are also being pursued experimen-
tally as a prosthesis for vision (for an up-to-date review see Luo and da Cruz,
90 Microneuronal Network Theories
FIGURE 3.7 A prosthetic MEA consisting of 60 electrodes used for electrical stimula-
tion of the retina.
Reproduced from Luo and da Cruz (2014) with permission.
–5
–20
N1 N2
–5
+15 +10
N3 N4
+10
FIGURE 3.8 A simplified “toy” network consisting of four neurons, N(1-4). Arrows
indicate connections between neurons and the numerals indicate the strength of
these connections. Although relatively simple, this network is probably not amenable
to prediction and solution.
connections. The valence and magnitude of the connections must also be rep-
resented by appropriate values if the network is to be fully characterized. Neurons
not directly interacting would simply not be connected by lines, although this
stratagem might obscure indirect routing of signals that would produce correlated
activity among even very remote neurons in the network.
It is also generally assumed that such a map would have to indicate in some
way that the constituent neurons would be dynamic, that is, capable of being
in different states at different times (otherwise the network would not be func-
tional, just locked in a constant state). The momentary state and the sequence
of states are additional properties that have to be specified if a network theory
of cognition is to be complete. Figure 3.8 is an example of the neuronal network
for a highly simplified “toy network.”
In this figure, signed numerical values indicate the potency of the intercon-
nections between neurons (negative values being inhibitory and positive vales
being excitatory) and the arrows indicate the direction of interconnections. At
each moment in time, we assume that the ensemble of states of all of the neurons
(Nn) in this hypothetical “toy” network would determine the network’s function;
at other times and other states of the component neurons, different functions
would be encoded by different configurations of this same network.
Extrapolating this basic idea of a “toy” map to the brain would constitute a
neuroreductionist theory in which the momentary states and interconnections
of the network would presumably be associated with a particular mental process.
Despite some limited progress with simple invertebrate nervous systems (where
real networks contain a small number of neurons), extrapolation of the network
concept to realistic vertebrate neural networks has not proven to be possible
currently. The best we can do is speculate about general properties (e.g., neurons
change their state as a result of changing synaptic efficacy).
It is also important to remember that the behavior of even as simple a toy
network as the four-neuron example shown in Figure 3.8 may itself be unpre-
dictable. Indeed, it is well known in mathematical circles that even a three-neuron
Microneuronal Network Theories 93
network may create a situation that is the formal equivalent of the three-body
problem in physics2—general solutions to which have long been known to be
intractable under the conditions of feedback and interaction that are characteristic
of all kinds of networks—biological and physical.
If even approximate, partial versions of cognitively relevant network maps are to
remain the immediate goal, then it is necessary to answer a number of initial ques-
tions concerning the data obtained with the MEA technique by means of appropriate
analytical procedures. The first question is—which neurons are members of the
network being examined? Of necessity, this task is complicated by a number of
issues. Trying to carry out such a task by repetitive probing with a single micro-
electrode would be tedious and ultimately unsuccessful because of the uncertainty
that successive penetrations would make contact with the same neuron. Such an
approach also ignores the loss of vitality of the neurons over extended periods of
time. The great potential advantage of the MEA technique with its ability to
simultaneously record from many neurons is that it promises to answer this first
question directly, if not completely. However, just how far one can go with even
1,000 electrodes in identifying a set of interacting neurons remains problematic.
Next, consider that the extracellular potentials picked up with an MEA are
not simply associated on a one-to-one basis with a particular electrode. Not
only will the activity of individual neurons be detected by more than one elec-
trode, but also one electrode will pick up the activity of several neurons. This
raises the second basic analytical question—which of the recorded responses go
with which of the identified neurons? This task is known variously as “response
sorting” or “clustering”—the assignment of responses that may have been picked
up by multiple electrodes to their source neurons. This may require the parsing
and then assignment of responses from among the noisy mixture of responses
recorded by each electrode. (See Figure 3.2B for a realistic example of responses
picked up from an array of electrodes.)
The third analytical question that has to be answered is what are the specific
functional interconnections and neuron states of the putative network that cor-
respond in what is still a mysterious way to some cognitive or mental activity?
This search for the neuronal network configuration or code is the “holy grail”
of cognitive neuroscience. The complete specification of the relationships between
mind and brain at this level would constitute fulfillment of the quest for a solu-
tion to the mind–brain problem. Unfortunately, for practical and conceptual
reasons we now consider, this “grail” may remain as remote and unobtainable
as the mythical one that Arthur’s knights failed to find.
analysis quickly fail if the property is neither linear nor stable. Pillow et al.
(2013) have been especially critical in noting the deleterious effect of overlap-
ping SAPs on spike clustering techniques.
6. Simple mechanical and crowding problems of the ancillary signal recording
and transmission equipment are a continuing problem. Neurons are micro-
scopic entities; throughout the history of microelectrode use, both single and
when enhanced by MEAs, investigators have had to fight the battle of making
their instruments small enough so that they do not interfere with the target
neuron mechanically. Furthermore, just getting the necessary parts of the
recording system into the container holding the neuronal tissue remains a
continuing engineering challenge.
7. Similarly, any nonrigid electrode configuration that is free to move about can
lead to what Lewicki referred to as “electrode drift”—gradual changes in the
spatial relationships of neurons and electrodes. This will inevitably lead to
changes in the shape of the recorded responses and inhibit SAP sorting and
clustering. Stabilization of electrodes has been greatly helped by the applica-
tion of photolithographic techniques from computer fabrication engineering.
Indeed, in most in vitro experiments, printed electrodes are perfectly stabilized
in the containment well in which the (typically dissociated) neural tissue is
placed. Unfortunately, because of tissue flexibility, stabilizing the relative posi-
tion of the neurons and the electrodes either in vivo or in vitro experiments
is still an unsolved, but serious, problem. Thus, the spatial arrangement of the
interneuronal network is likely to fluctuate. Although some progress has been
made in “caging” single neurons near single electrodes (e.g., Erickson, Tooker,
Tai, and Pine, 2008), the problem remains a real one for those who wish to
study neuronal populations over prolonged periods of time.
8. Lewicki also pointed out that with any communication system, noise is
always recorded superimposed upon the signal information of main interest.
Depending on the signal-to-noise ratio, critical information may be fully lost
if this ratio is too small. Even worse, if the noise is not stable, changing prop-
erties from moment to moment would make the analysis process especially
challenging. Indeed, unstable noise may completely obliterate any possibility
of sorting SAPs using any of the measures mentioned here.
(Paraphrased from Lewicki, 1998)
of the number of neurons and synaptic connections in neural tissue, but this
misses the whole point.
Network complexity, therefore, is viewed as a property of systems with numer-
ous parts and inter-part connections. Measures such as the factorial can provide
estimates of “complexity,” but, in general, the utility of these numbers collapses
as the number of components and interactions increases and numbers such as
the factorial soar to hypercosmological levels. For computer scientists, the com-
plexity of a problem has to be computable in order to lead to criteria for the
solvability or convergence of certain algorithms. For cognitive neuroscientists
operating at the microneuronal network level, almost everything beyond three
neurons is, for all practical purposes, “infinitely” complex.
The reason for this uncertainty forewarns us about one of the most funda-
mental limitations of any technique that seeks to provide an effective map of a
neuronal network. It is something that is likely to raise the hackles of those who
seek to represent the structure of a neuronal network but it is essential that those
who seek to use MEAs to disentangle neuronal networks understand its con-
notation, if not its denotation. Horwitz (2003) went on to say:
Should this be true for a pair of neurons, how much less likely is achieving
the goal of mapping the “functional interactivity” of a suitably sized collection
of cognitively significant neurons? Thus, the quest for a unique effective map
may always be a search for a chimaera; the best we will be able to do will be
to develop a functional description—effective maps being denied to us.
Despite these musings, work continues to determine the organization of
realistic neuronal networks—mostly with invertebrate model preparations. It is
worthwhile reiterating that although some progress has been made in response
clustering of simple preparations, the grand promise of producing a solution to
the general network problem (mapping or decoding the specific neuronal net-
work for a particular cognitive process) is still unfulfilled and nothing on the
horizon is promising.
An interesting place to start a discussion of this kind is with a work that can
only be described as a heroic example of a class project. This work is presented
here because of its pedagogical simplicity and conceptual clarity. It is unlikely
that this approach could prove to be an efficient procedure for mapping real
neuronal networks, but nowhere are the methodological and logical assumptions
behind neuronal network construction made clearer and the inevitable failure
of the “holy grail quest” highlighted.
100 Microneuronal Network Theories
Two students at MIT, Jonathan Karr and Daniel Herman, authored4 a labora-
tory semester project in which indications of microneuronal network connectivity
were determined by the following method. Their experiment5 used data recorded
from an in vitro culture of hippocampal neurons using an 8 × 8 MEA. Each
of the 64 electrodes was used in turn as a stimulator and the responses induced
in all of the other 63 electrodes measured. Correlations were then individually
computed between the stimulus applied to the single stimulator and the responses
measured at all of the other responding electrodes—a selection that might be
fewer than 63 because not all of the electrodes picked up a neuronal response.
This produced a maximum of 4,032 correlations (64 × 64 minus the 64 auto-
correlations of the stimulator with itself). The amplitudes of the correlations
were interpreted to be measures of the degree of interconnectivity among the
electrodes; high correlations, driven by high amplitudes and short latencies, being
tantamount to intimate causal interconnections and spatial propinquity between
the stimulating and the responding neurons.
Karr and Herman’s brief report is exceptional not only because of its successes
but also because of the way it highlighted some of the limitations of the whole
idea of using MEAs to determine the connectivity of a network of neurons.
On the positive side, their technique embodied the ability to provide initial
estimates of a number of the connectivity parameters of the hippocampal neu-
ronal network they were studying. Based on the assumption that the probability
of an inferred connection would be higher for a correlated pair of electrodes
than for an unstimulated pair and such other factors as the delay between the
stimulus and the response, preliminary estimates could, in principle, be made of
factors such as the geometry of connection (e.g., how far apart the stimulated
and the responding neurons were either spatially or in terms of intervening
neurons) and the interconnection strength (i.e., the magnitude of the correlation
between the single stimulating and the other 63 responding electrodes).
Despite the unusual accomplishments of these students, their work also dem-
onstrated the extreme difficulty of the task of determining the connectivity of
a neural network with MEAs and the most powerful analytical techniques avail-
able. For example, there is a cryptic assumption in experiments of this kind that
an ideal, pseudoregular neural network underlies these correlations. Neuroana-
tomically, however, the irregular geometry of a real neuronal network does not
even come close to that of the regular 8 × 8 array of the microelectrodes used
to collect these data. Collectively, the real neuronal networks are much more
irregular. Whereas a realistically complex neuronal network may have dozens, if
not hundreds, of synaptic connections between pairs of neurons and many
thousands on each neuron, the analytically derived network implicitly assumes
a much simpler kind of network based on a much simpler concept of connec-
tivity than is actually present.
To exacerbate this situation even further, as I noted earlier, each extracellular
SAP may be detected by more than one electrode and that one electrode may
Microneuronal Network Theories 101
FIGURE 3.9 The effect of arbitrary thresholds on the connectivity pattern of a neural
network.
From Karr and Herman (2004) with permission.
record activity from many neurons. Therefore, even under the best conditions,
analyses with even the most limited goals are constrained to deal with small
numbers of neurons and unverifiable ambiguities.
Another issue is the arbitrariness of the interconnectivity networks generated
by methods like those reported by Karr and Herman. The main measure used
in their analysis—correlations—as a measure of interconnectivity can be instanti-
ated in a large number of different maps depending on arbitrary decisions. If a
low threshold for a correlation is used as a criterion for interconnectivity, the
resulting pattern would suggest a large number of putative links. On the other
hand, if a high threshold was used, only a few links between neurons would be
generated, as shown in Figure 3.9. The essential problem is the arbitrariness of
the threshold for what would be acceptable as a connection.
Nevertheless, the importance of Karr and Herman’s laboratory exercise went
far beyond the preliminary nature of their accomplishment. Others (e.g., Maccione
et al., 2012) also subsequently reported the use of correlational techniques similar
to those used by these two students. Whereas Karr and Herman stimulated the
electrodes to define correlated responses, Maccione and his colleagues allowed
spontaneous activity to define causal functional relationships and thus to imply
interneuronal connections. In addition, Maccione et al. also used filtering techniques
to reduce the number of spurious connections—combinations of distance and
latency being used to eliminate at least some of the implausible connections.
In recent years a number of other more sophisticated schemes have been devel-
oped that purport to determine the connectivity pattern of a neuronal network
by means of MEAs and novel analytical techniques. One of the most comprehensive
is the work of Stevenson et al. (2009). This group applied Bayesian methods to
determine the statistical properties of neuronal networks in both toy and in vivo
preparations of monkey brain neurons. Their goal was to characterize the functional
connectivity of networks in the form of a matrix in which causal relations between
pairs of neurons were either present or not. Their data were presented in the form
of a matrix that showed which neurons were interacting with which others. Scores
were binary values of interactions between all possible pairs of neurons in a man-
ner comparable to the correlational scores reported by Karr and Herman. However,
neither group of researchers was able to produce a connectivity map of the general
type desired nor could they distinguish which interconnections were monosynaptic
and which were the result of indirect pathways.
102 Microneuronal Network Theories
To accomplish what they did, Stevenson and his colleagues had to make
assumptions about the nature of the neuronal networks they were studying.
These simplifying assumptions (“priors” in Bayesian terms) made it possible to
avoid a combinatorial explosion, but also limited how detailed the inferred
networks could be. Their priors included, first, the assumption that neuronal
interconnections are sparse, a generalization that almost everyone agrees is not
anatomically true. Second, they also assumed that neuronal interactions are tem-
porally smooth, that is, that the interactions occur without abrupt discontinuities
in time. This assumption is also probably incorrect.
In summary, we now list some of the difficulties and challenges in reported
attempts to determine the connectivity of a neuronal network:
1. At the base of many of the other problems faced by those who attempt to
map out a realistic neuronal network is the fundamental underdetermina-
tion of the data obtained with an MEA. That is, it cannot be definitively
determined by any analytical method yet proposed which of several (many?)
equally plausible network configurations is responsible for a functional
response. Neuron X may be connected to Neuron Y by a single direct syn-
apse or by an indirect pathway that may involve a number of neurons. Any
pair of neurons may be interconnected by a multitude of alternative network
connections. Simply demonstrating a correlation does not adequately define
a specific anatomical connection. Indeed, there is a strong possibility that it
cannot, in principle, do so.
2. There is a discrepancy between neuronal networks determined electrophysi-
ologically and those produced by analytical methods. It is often necessary
to assume that there is a sparse pattern of connectivity (often approaching
one-to-one connections) to avoid generating intractable combinatorial prob-
lems. In anatomical fact, however, all of our microscopic evidence suggests
that each neuron is complexly and multiply interconnected with many other
neurons, by as many as 1,000 synapses.
3. Unlike the neat grid-like arrangements of MEAs, neurons are irregular and
idiosyncratic in both their shape and connectivity.
4. In many cases, real neuronal networks are influenced by neuronal activity
from outside the network under study. Thus, both centrifugal and centripetal
sensory or motor activity may be influencing what was originally thought to
be a network that could be studied in isolation.
5. Almost all of the current methods are applied in experiments that, of neces-
sity, report the use of relatively small numbers of neurons in their analyses.
The availability of the MEA suggests, erroneously I believe, that this is no
longer an issue since the activity of large numbers of neurons can be recorded
simultaneously. However, a careful scrutiny of the literature suggests that the
interconnectivity pattern of only a few neurons has actually been determined.
In short, 1,000 electrodes in an MEA does not mean that the interactions
Microneuronal Network Theories 103
Similar calculations can be made to show that there are many other problems
that cannot be solved in the lifetime of the universe. Therefore, neither time
104 Microneuronal Network Theories
nor space is adequate to deal with the complexity of even relatively (superficially)
simple problems in the lifetime of the universe.
Karp (1986), a prominent complexity theorist, also reported how many other
superficially simple combinatorial problems, many of which were analogs of
those faced in cognitive neuroscience, were actually known to be intractable in
their demands on time and space. Karp’s (1986) work on the intractability of
many other superficially simple combinatorial problems is especially relevant in
this context. Specifically, he pointed out that many, if not most, real-world
problems are said to be NP-complete, that is, they cannot be solved in real
amounts of time. Although the work of Karp and others is well known in
mathematics and computer theory, cognitive neuroscientists seem to be assuming
that some of their greatest challenges are in principle solvable when it has already
been proven that they are intractable.
Despite these reminders, there has been an implicit acceptance of the idea
that the mind–brain problem will ultimately be resolvable at the microneuronal
network level. (See the discussion in Chapter 5.) In this section, I consider some
basic arithmetical background suggesting that any attempt to develop a con-
nectionist model of anything approaching a cognitively significant neuronal
network is not a feasible experimental goal.
By far the most profound challenge facing those who seek to determine the
functional topology of real neuronal networks is what can be designated as
“combinatorial or computational complexity.” The general term “complexity”—
a phrase thrown about by cognitive neurophysiologists with total abandon and
freedom from any precise definition of the term—is heard in virtually every
conversation about the brain. The term, regardless of how deeply we understand
its numerical implications, certainly has profoundly affected our research by both
explicitly and implicitly limiting our research efforts to toy networks with only
a few neurons or to macroneural networks.
The constraint that inhibits progress in understanding network theories of all
kinds is how large the number of alternative organizations of even a very small
number of interconnected components can be. The basic arithmetical terms that
are germane to the topic of complexity are combinations, factorials, and permuta-
tions. These terms are relevant to neuronal networks because the number of
alternative, yet equally plausible, organizations of a very few components or nodes
can very quickly rise to huge numbers.
Factorials (n!) are particularly relevant to neural networks because they cor-
respond to the number of different ways that n objects (e.g., neurons) can be
arranged (e.g., interconnected) into different groupings. The sum of all of the
possible maps that are necessary to fully characterize a network quickly becomes
enormous. The factorial n! of a number n is defined as the multiplicative product
of all of the positive integers that are less than or equal to n. Even for small
numbers, surprisingly large factorials are generated. For example, although 3! is
equal to 6, 10! is equal to 3,628,800. Larger factorials (for example, 25! is equal
Microneuronal Network Theories 105
to 1.551121004 × 1025) very quickly approximate and then exceed the number
of basic particles in the universe. This numerical explosion provides a good idea
of what happens even when one studies relatively simple networks.
Combination (C) is a more specific term that is defined in terms of the
number of ways that a particular subset of objects can be arranged. The formula
for combinations of n things taken r at a time is:
n!
C=
(n − r )!(r !) Eqn. 3.1
C can also get very large very quickly as n grows. Consider the classical example
of playing ordinary 5-card poker for which n = 52 and r = 5. Applying Eqn. 3.1
specifies that there are 2,598,960 such combinations if one ignores the order in
which the items are selected. Considering how much larger n and r can be in
the brain, the number of coded neuronal combinations is also enormous.
If the order of the ways that a set of things are combined matters, then the
word “combination” has a more restricted meaning—it becomes what is called
a permutation. Because the particular way in which a set of neurons is function-
ally arranged would presumably lead to different cognitive processes, it is probably
more correct to speak about changes in brain states as permutations rather than
combinations.
Furthermore, this is not the worst case. There are other mathematical series
that grow even faster than factorials or combinations—e.g., superfactorials—that
although not a priori irrelevant to this discussion help to give us an intuition
about brain complexity. A superfactorial of n is the product of the first n fac-
torials. All such “super” functions make the same point: exhaustively studying
networks or seeking to identify the unique arrangement that encodes a cognitive
process is probably not currently a feasible or plausible approach to generally
solving the riddle of how the brain makes the mind. Although we may be able
to make some statements about the general properties of such networks, it is
obviously impractical to attempt exhaustive experiments to understand their
function. This idea of factorials as measures of possible organizations exceeds by
far the already serious problems generated by the simple numerousness of the
neurons in the brain.
The link of these esoteric mathematical concepts to neuronal networks is
clear. Even if we were able to represent the brain as a system of billions of
uniform and simply interconnected “billiard balls” it would underestimate the
true complexity of the brain. Neurons are not simple entities as are billiard balls,
but are themselves complex entities with many idiosyncratic properties and a
variety of interconnections of varying weight.
Koch (2012) recently put this problem into perspective when he calculated
that the time required to “exhaustively describe a [neural] system” (and this is
what we assume we have to do to explain how the brain generates the mind)
106 Microneuronal Network Theories
would grow faster than exponentially. As an example, Koch calculated that the
2 million neurons of the visual cortex of a mouse would require a computer
system running for 10 million years even if the computer’s powers were simul-
taneously growing at the modern rate suggested by Moore’s Law (“computer
component density and thus computational speed doubles roughly every 2 years”).
Koch did go on to note that there were some simplifying conditions (compa-
rable to those used to speed up a standard Fourier transform to a “fast” FFT
version) that could speed up the process by a few orders of magnitude. One
way to do this is by grouping neurons into clusters. However, whenever one
regularizes a procedure in this manner, there is the ever-present possibility that
the cognitively relevant microneuronal organizational properties one is seeking
may be tossed out.
Koch (2012) also reminded us of another measure of complexity—Bell’s
number—defined as the number of different partitions into which a system
of n objects can be divided. For example Bell’s number for n = 10 is 115,975.
Playing out its calculations serves as another reminder of the complex ways
in which the components of a network can interact or be divided. A Bell’s
number of 115,975 compares to factorial 10 which is equal to 3,628,800.
Thus, factorials grow faster than Bell’s numbers (and both scale faster than
exponentials) but the effect is the same—huge eventual demands for analytical
speed if we are ever to exhaustively study the brain as an example of a neu-
ronal network.
Whatever the best measure, they all pale into insignificance compared with
the actual number of neurons in the human brain, and much of current think-
ing is that almost all of them participate in encoding cognition. The only
hope, according to Koch, of ameliorating this constraint of “colossal” numbers
of calculations is to find some way to reduce the number of subunits so that
the number of possible partitions can be reduced. However, the irregular nature
of most brain tissue (notwithstanding the discovery of such prototypes as
repetitive columnar organization) makes it likely that we will join such alter-
native mathematical treatments as “graph theory” in attending not to the
specific problems of cognitively realistic networks but rather to the general
properties such as reachability, directedness, and degree of toy networks. As
interesting as the work on graph theory of Harary (1969) or Christofides
(1975) may be, little attention is paid by graph theorists to predicting the
function of complex graphs or the differential effects of different partitions.
Their attention is drawn to general properties of very simple graphs that may
extrapolate in some ill-defined way to larger networks. This decision by graph
theorists to concentrate on simplified abstract models is well chosen since, as
noted earlier, even the three-body (or, in this case, the three-neuron problem)
is probably not generally solvable.
“Complexity” has other ramifications beyond simply the raw numerousness
of the nodes and connections on which I have concentrated so far in this
Microneuronal Network Theories 107
There is also no other organ system for which the complexity of the
structure is so great that earnest arguments can be made against delving
into structural analysis because such an effort might well provide an
unprecedentedly gigantic, yet totally incomprehensible, mass of data.
(p. 618)
Although Lichtman and Denk express optimism that the task of mapping
neuronal circuits may be tractable sometime in the future, as of this date I have
not been able to find any report of a cognitive (i.e., something beyond a periph-
eral motor or sensory mechanism) linkage to vertebrate microneuronal networks
that meets the criteria expressed by Martin, Grimwood, and Morris for robust
empirical proof at the microneuronal level.
108 Microneuronal Network Theories
may display the same shapes and amplitudes and the same neuron may be
characteristically different from one trial to the next. No single discriminat-
ing criterion for solving the clustering problem works in a general way.
5. Recording is fortuitous: Some electrodes will record nothing and some will
pick up the responses from multiple neurons in an unpredictable manner.
6. The regular pattern of an array of 100 such microelectrodes (e.g., as shown
in Figure 3.2B) does not necessarily correspond to the spatial arrangement
of the neurons in the neural tissue under investigation. That is, beyond the
sensory and motor representations, they may not be topologically consistent.
Therefore, it may not be possible to map directly from electrode recordings to
the irregular pattern of a cognitively significant neuronal network.
7. The irregular distribution of synaptic connection also contributes to the lack
of repeatability of experiments.
8. The biological and chemical responses to the foreign matter from which the
electrodes are constructed can mitigate their use in long-term applications—a
requirement for clinical use. Similarly, long-term implantation of electrodes
remains a major challenge. A good discussion of the difficulties encountered
when one attempts to record or stimulate for long periods can be found in
the review by Moxon et al. (2009).
9. Even the most minute array of electrodes will do damage to neurons and
interconnections. Efforts to develop flexible and compliant microelectrode
arrays have met with only moderate success.
10. Much more devastating to the hope that the organization of realistic neural
networks will be uncovered using MEAs is the enormous amount of data that
is obtained from even a relatively few irregularly spaced neurons. If the goal
is to understand how multiple neurons interact, then the persistent problem
of combinatorial computability produced by the complexity of neuronal net-
works cannot be overlooked.
The thesis of this chapter is that partly because of technical and procedural
limitations, but mostly because of the intrinsic nature (i.e., the complexity) of the
problem, the direct, exhaustive, brute force strategy of stimulating and recording
the individual components of a microneuronal network that worked so well at the
level of single neurons is not likely to ever be implemented at the microneuronal
level for ensemble processes. Microneuronal network theory building in such a
many-neuron environment is severely inhibited because of the improbability of
ever simultaneously and controllably stimulating and measuring the activity of the
number of neurons involved in even a “simple” cognitive process. Thus, the
possibility of developing an authentic, valid, overarching microneuronal theory of
cognition based on robust empirical evidence appears to be remote. This is so
despite the fact that there are no plausible alternatives to Hebb’s conjecture and the
speculative theories that flow from it. Hebb was right, most cognitive neuroscientists
seem to agree, but there is no way to prove it in detail. Although we may all agree
110 Microneuronal Network Theories
Theories Technologies
Mind is in the heart Dissection and trauma
Mind is in the head Behavioral observations
Mind is distributed EEG
Mind is localized fMRI
Single neuron theory Microelectrodes
Neuronal networks Computers and MEAs
Hebb’s conjecture Plausible speculation
Source unknown.
Notes
1. The optogenetic technique is a major step forward in neurophysiological research. By
various methods (such as using a virus that carries a light-sensitive trait) genetic material
can be introduced into a neuron (or, more usually, a class of neurons) that makes it
respond differentially to light stimuli of different wavelengths or temporal patterns.
Manipulations of opsins, the light-sensitive materials of the eye, are particularly useful in
this regard since their optical properties are well known. Because of the different optical
sensitivities of the genetically manipulated materials, specific neuronal or synaptic types
can be turned on or off by stimulating the neurons with different kinds of light. In the
Nabavi et al. study the virus used to vary the behavior of the synapses produced neurons
that were selectively sensitive to low-frequency trains of light. Specifically, the optical
112 Microneuronal Network Theories
stimulus–producing LTD was 900 pulses of light, each 2 msec, at 1HZ (Nabavi et al.,
2014). The stimuli-producing LTP was 5 trains of light (each train 100 pulses, 100Hz.
In this manner, the degree of synaptic activity could be manipulated and direct control
over their function achieved for the first time.
2. The three-body problem has been most extensively studied in physics in the context
of planetary motion. It has been known since the 19th century that the behavior of
three gravitational interacting bodies cannot be solved in general unless one of the
bodies is much smaller than the other two. The dynamic behavior of three equal-sized
objects is, as far as we know, nonrepetitive and unpredictable.
3. Sporns’ (2011) book is an outstanding discussion of the nature of both anatomical
connectivity and connectivity in general. It is a necessary part of the education of any
cognitive neuroscientist. It tells us much about networks, but it says very little about
the relation between the anatomical connectome and cognition.
4. Unfortunately, this work seems not to have been published (it is dated March 31, 2004);
more complete details can be found in a single entry to it here: http://web.mit.edu/9.29/
www/neville_jen/jkarr/mea/. It was, however, so interesting in the present context that
I could not ignore it.
5. It was not entirely clear from their report whether they collected the data they analyzed
or whether prerecorded data were provided to them for analysis. In an April 2014
personal communication from Karr, he indicated to me that they used prepackaged
data as the raw material for their analysis. For the present discussion, it does not really
matter since the goal of their paper (and of the present discussion) was to infer the
connectivity of the microneuronal network described by the data. Having prepackaged
data made the class project possible.
6. Of course, this number will change depending on whatever computer is available at
any given time. However, the numbers are such that the Stockmeyer theorem argues
that no conceivable computer could perform the necessary calculations in the lifetime
of the universe. Whatever the specific numbers, this theorem does provide an intuitive
foundation for the gap between computational needs and computational capabilities.
4
INTERMEDIATE LEVEL
NEURONAL NETWORKS—
THE ROLE OF MODELS
4.1 Introduction
A major goal of modern cognitive neurophysiology is to establish connections
between cognitive processes and the pattern of neural activity embodied in
networks of neurons. Obviously, however, this is going to be an extremely chal-
lenging goal for systems as complex as the human brain where innumerable
neurons are presumably involved in the simplest of thoughts. When dealing with
complex networks potentially involving millions and billions of neurons, the
technical and conceptual problems inhibiting this kind of research are profound
as was pointed out in the previous chapter. For all practical purposes, combina-
torics and numerousness probably will not permit experiments of the necessary
theoretical kind—for the present, at least, no empirical neurophysiological studies
of the detailed arrangement of cognitively realistic neuronal networks have yet
been carried out. An alternative strategy, of increasing importance in cognitive
neuroscience, is the use of relatively simple, naturally occurring model systems
that mimic or model some aspect of the behavior of much more complicated
organisms—hopefully with a simple analyzable neural network. The immediate
goal of this line of research is to at least understand the basic principles that
may help us understand how very complex neural networks operate by studying
the accessible and tractable neural networks of a simple creature.
Lying between the extremes of single neurons and full-blown neuronal net-
works, however, is an intermediary region in which there is a body of empirical
neurophysiological evidence and for which the networks are simple enough to
permit at least approximate descriptions of their functionality. This approach
depends on the discovery of relatively simple neuronal networks in “model”
organisms that are capable of mimicking what are almost certainly more complex
114 Intermediate Level Neuronal Networks
8. It also helps if the nervous system of the model is stable from specimen to
specimen so that the involved neurons can be individually identified and
experiments be repeated.
9. Because many experimental procedures are now based on genetic manipula-
tions, having a more complete knowledge of the model animal’s genome is
also advantageous.
10. The design of an experiment must be able to distinguish between a necessary
and a sufficient theory of network organization (Olsen and Wilson, 2008).
11. Analogous behaviors must be carried out by both the modeled and the model
system. “Analogous” in this case means that the behavior accomplishes the
same function (such as learning) in both the modeled and the model organ-
isms. It does not mean that exactly the same kinds of networks are used to
implement both. Indeed, we typically do not know what mechanisms exist in
the modeled organism to produce the behavior—that is why we have turned
to simpler models for enlightenment.
12. The behavior of the model organism must be measurable and the network
configurations controllable in order to carry out controlled experiments. A
variety of methods, including genetic and neurophysiological techniques, are
now available to control behavior.
13. The a priori judgment that the structure of the model is exactly the same
as the structure of the modeled network is almost never provable. The rela-
tionship between the underlying mechanisms of each is one of analogy, not
homology, and though behaviorally similar, there remains uncertainty about
the salient networks of the model and the modeled, respectively.
Physical
Subjective
Distance
FIGURE 4.1 The shape of the physical and subjective responses to a gradient of
illumination.
Reproduced from Uttal (1973).
Intermediate Level Neuronal Networks 119
by reciprocal interactions between different points on the retina and/or the rest
of the perceptual nervous system. Of course, in the 19th century, it was not pos-
sible for researchers to carry out the necessary experiments to confirm this
hypothesis, but in the latter half of the 20th century, it became eminently feasible,
if not yet on the vertebrate visual system, then on a simpler model preparation—
the visual system of Limulus polyphemus—the North American horseshoe crab.
The horseshoe crab is a vestige of organic evolution that has survived from
roughly 400 million years ago. It is an arthropod but actually unrelated to crabs
and other crustaceans; it is more closely related to spiders and other arachnids.
Fortunately for cognitive neuroscience it is also graced, like spiders, with multiple
eyes on the top of its carapace. Two large compound eyes, in particular, are of
special interest because each consists of approximately 1,024 relatively large cell
complexes called ommatidia. Each ommatidium contains a number of secondary
cells, but lateral neuronal connectives from the eccentric cell (the main light-
sensitive neuron in the ommatidium) are the main links to other cells of the
complex ommatidium. By this means, light-induced neural responses interact
with surrounding ommatidia through lateral connections that can be seen in
optical microscopes (as shown in Figure 4.2—a drawing of the neural network
in one ommatidium).
ret. c. Pigment
rhab.
C
Eccentric cell
Neuropile
A sh. cell nuc.
The relatively huge size of the cells in the ommatidia first led researchers
(Hartline and Graham, 1932) to use them as a model preparation for the study
of single neurons. However, an inspired epiphany stimulated the group led by
Hartline (1940) to consider the L. polyphemus eye as a model of interneuronal
interactions. The basic paradigm that Hartline and his coworkers used was to
stimulate a narrow region of the crab’s eye with a spot of light, measure its
neural activity with electrodes, and then stimulate it with another spot in a dif-
ferent location. The typical result was that the addition of the second light
stimulus would reduce the response from the ommatidium from which the
recording was made. Although only a single electrode was used to make this
measurement (and the response of only a single ommatidium recorded at a time),
both the electrode and the two spots of stimulating light could be moved across
the compound eye to track out the spatial pattern of the response as a function
of such variables as the distance between the two light spots. By this reiterative
procedure, it was thus possible to deal with this model preparation as if multiple
electrodes were recording from a spatially distributed response to a spatially
distributed stimulus—sequential recordings taking the place of a simultaneous
spatial distribution of responses.
This technique may have been the first experiment to deal neurophysiologi-
cally with neurons interacting in a microscopically defined real neuronal network.
Previously only recordings had been made from single neurons. The work carried
out by the Hartline group was, therefore, a major milestone in the study of
neuronal network theory thanks to the unusual nature of an evolutionary
anomaly—L. polyphemus. Although the number of neurons in the creature’s
eye was larger than what I have designated here as an intermediate level, the
extreme regularity of the anatomy made it possible to deal experimentally with
the network properties of this unusual eye. It is problematic whether this feat
has been repeated in any other model preparation.
The specific question asked in this pioneering series of experiments was—does
the presence of the second stimulus affect the first region’s response (Hartline,
1949)? The answer to this questions was that yes, it certainty does and in a very
specific manner whose properties could be measured—the sine qua non of a
model preparation. The interactions between ommatidia were solely inhibitory,
diminished in strength with the distance between stimulated regions, and were
also mutual (Hartline, Wagner, and Ratliff, 1956)—the effect of one on another
was the same as the other’s on the first. The main implication of this research
was to open the door to the study of the simultaneous interactions of many
neurons in the form of a network. It provided a compelling explanation of the
Mach Band phenomenon by demonstrating a neurophysiological analog of edge
enhancement, thus validating what had only been an inspired guess by Mach.
This simple model invertebrate system thus also provided a compelling
explanation of the similar neural interactions that presumably were going on in
the human visual system. It was determined that these lateral inhibitory
Intermediate Level Neuronal Networks 121
This does not mean that the final explanation of this phenomenon in humans
is at hand. That the neurophysiological properties underlying edge enhancement
in the L. polyphemus and the human perceptual phenomenon are mediated by
exactly the same simple neuronal networks seems intuitively unlikely. The Hartline-
Ratliff “theory” that the neural and perceptual phenomena are alike in principle
is based on a very simple sensory mechanism for which there is ample neuro-
physiological supporting data and in which the neuronal network is relatively
simple in structure. In other words, although the L. polyphemus eye effect is based
on local geometric interactions, we still have not established that this is also the
case in the human. Nevertheless, what we do know about the neurophysiology
and neuroanatomy of L. polyphemus is probably still the best model we have
today in which the neural and psychophysical data are so mutually consistent.
In sum, the Hartline et al. research program was one of the first, if not the
first, to develop a mathematical theory of observed neurophysiological network
mechanisms that might underlay a psychophysically reported illusion—a discrep-
ancy between the stimulus and the resulting experience based on the properties
of an interactive network. It did so by providing a body of relevant experimental
data that built on direct measurements of neurophysiological responses. The
model has been widely accepted because it reproduced the perceptual experience
in a simple model neurophysiological preparation. It is however, based on another
unproven assumption, namely that the spatial mapping of the neuronal response
in L. polyphemus is recapitulated in the perceptual experience of the human
observer. This gap in the logical chain remains unbridged.
The key to this theoretical success was the susceptibility to mathematical
modeling provided by the simplicity of the L. polyphemus eye. An important
aspect that makes this analysis so compelling is that the mechanisms used to
explain the overall system responses are from microneuronal interactions—that
is, network properties. The edge enhancement phenomenon would not be
observed in the activity of a single ommatidium. It was only when multiple
components were shown to interact with one another that the neuronal con-
nections and interactions of this compound eye exhibited the edge enhancement
phenomenon. The presumption is that similar (in principle) but probably much
more complex neuronal networks account for the Mach Band perceptual illusion
reported by humans. If nothing else, this work was one of the earliest to provide
an empirical foundation for a fundamental postulate of cognitive neuroscience—
that it is from the information-processing capabilities of neuronal networks that
a preliminary understanding of cognition is most likely to come.
L7
and stage L8
LD LD RCQG
R14
L10 R13
(c)
Siphon
Mantle shelf
Gill
INT
Polysynaptic
L7 LD L9 L9
G 1 2
SENSORY
N.
Monosynaptic
FIGURE 4.3 The anatomy of the gill and siphon neural mechanisms in Aplysia cali-
fornica. (a) and (b) show the gross anatomy of the system, and (c) shows the six
neurons involved in withdrawal from noxious stimuli, some of which are
redundant.
Reproduced from Kupfermann et al. (1970) with permission.
128 Intermediate Level Neuronal Networks
sample of C. elegans to the next, but also each neuron has been identified and
both its shape and its location mapped. According to Sporns (2011), this is the
only model creature for which such a complete nervous system map currently
exists. The mapping was an enormous task carried out by White, Southgate,
Thomson, and Brenner (1986) based on an exhaustive reconstruction of a series
of electron micrographs. As a result of this complete mapping and some inge-
nious research techniques, C. elegans has proven to be a useful model animal
for studying the relations between behavioral and neurophysiological
processes.
However, this approach has not been without difficulties and challenges. Why
should this be the case given the detailed nature of all of the neurons in the
brain map we currently possess? The main answer to this question is that the
C. elegans nervous system, despite its small size, is considerably more complex
than the other two model nervous systems just discussed. The word “complex-
ity,” so easy to articulate and so difficult to explain, is an expression of a general
principle permeating this entire book—the complexity and numerousness of the
neuronal network of even these simple model animals pose combinatorial prob-
lems that quickly explode to demand unachievable computational powers.
The situation concerning the use of C. elegans as a model nervous system,
therefore, is quite different than the ones characterizing A. californica or L.
polyphemus, both of which have simplifying attributes. In the former, it was pos-
sible to distill the germane interactive network function down to a single repeti-
tive functional process (mutual lateral inhibitory interaction) in which the
arrangements of individual nodes, the ommatidia, were almost crystalline in their
regularity. This made iterative computational algorithms simple enough to be
executed. In A. californica, it turned out that the germane functional process
(habituation) was actually ascribable to mechanisms that could be localized to a
single interneuronal junction.
Furthermore, both A. californica and L. polyphemus are relatively large animals
compared with C. elegans. Thus, they could be manipulated and probed with
stimulating and recording instruments in a way that the smaller neurons of C.
elegans could not. The ommatidia of L. polyphemus and the axons and ganglia
of A. californica were large enough to be easily manipulated in an electrophysi-
ological experiment. The small size of the animal made simple mechanical
manipulation of C. elegans difficult—special techniques had to be developed to
immobilize the creature’s miniature worm-like body. C. elegans neurons were
not only small but had the disconcerting property of bursting when punctured
with a microelectrode. Nevertheless, the task of electrophysiological recording
from this tiny creature was accomplished for neuromuscular junctions by Raizen
and Avery (1994) and then for neurons by Lockery and Goodman (1998) and
Goodman, Hall, Avery, and Lockery (1998). A useful compendium of practical
advice detailing how neurophysiological recordings can be made from C. elegans
can be found in an article by Richmond (2006).
132 Intermediate Level Neuronal Networks
Despite its small size, the resulting practical problems, and the relatively large
number of neurons in this “simple” nervous system, C. elegans has been success-
fully used as a model for olfactory aversive learning. Ha et al. (2010) were able
to compare changes in behavior as a result of experience with the operation of
a relatively small neuronal network consisting of 13 neurons located in a ganglion
in the anterior end of this nematode. The experimental procedure they used
was based on an innate behavior exhibited by C. elegans; this tiny nematode
responds to the activation of chemosensitive neurons in its head region to a large
number of chemicals secreted by the bacteria that form its food (Bargmann,
Hartwieg, and Horvitz, 1993). Interestingly, although many of these chemicals
are detrimental to the animal, instigating diseases or metabolic failures, naïve
animals that have not been exposed to these chemicals are attracted to rather
than repelled by the chemical odorants. It is only after some experience with
the negative effects of these chemicals that C. elegans avoids them. The task faced
by Ha and his group was to determine how and where this model nervous
system was organized to alter this behavior from attracted to aversive.
One of the first problems encountered was to measure the change in behavior
from the naïve (attractive) to the mature (aversive) state. Fortuitously, the pattern
of movement executed by C. elegans toward attractive chemicals differed from
its pattern of movement away from aversive chemicals. There was a much smaller
proportion of a kind of behavior Ha and his colleagues designated as Ω turns
(“sharp turns in which the animal’s body shape resembles the Greek letter omega
Ω”), when the animal behaved as if there was an attractive odorant present, than
when the behavior suggested that the chemical was aversive. Populations of C.
elegans could then be trained to be used as controls for those that had been
“taught” to avoid detrimental odorants and those that had not been. The pro-
portion of Ω turns during attractive behavior could then be compared with the
proportion during aversive behavior to determine whether the chemical was
attracting or repelling C. elegans.
To explore the questions of the neural correlates of this learned behavior
(from attraction to avoidance), it was then necessary to manipulate the C. elegans
nervous system to determine the role of individual neurons in the learning
process. This was accomplished by what was also an ingenious technique. Ha
and his colleagues used a laser whose beam size had been demagnified to the
point that it could target individual neurons. The purpose of the laser was to
destroy targeted neurons. Therefore, any changes in C. elegans’ behavior could
be attributed to the absence of that particular neuron. The results of their
experiments with this kind of selective destruction allowed them to map out
the details of a network of 13 neurons that accounted for the naïve attractive-
ness of the odorants, the learned avoidance of them, and the experienced avoid-
ance of the odorants. This analysis allowed Ha and his colleagues to uncover
the fact that two different configurations of a relatively simple neuronal network
Intermediate Level Neuronal Networks 133
Synapses:
AWB AWC
1 - 10
11 - 20
AIY
ADP 21 - 30
31 +
AIZ
Gap junctions:
RIA AIB
1-3
4+
mediated the naïve attractiveness and the trained aversion responses to the odor-
ants. The network of neurons mediating the two forms of behavior as well as
the learning process that changed the behavior is shown in Figure 4.4. Each of
the blocks identifies one of the 13 neurons that are interconnected by the number
of synapses indicated in the side bar.
Ha and his colleagues summarized the role of destroying various neurons on
the naïve (attractive) and learned (aversive) chemosensitive behavior as well as
the role individual neurons had on the learning process itself. With the pattern
of observed neuronal interventions and behavioral changes they were able to
construct the neuronal network shown in Figure 4.4—a remarkable tour de
force made possible by the evolutionary accident of C. elegans’ anatomy. The
following list abstracts a few of the conclusions they drew from their experi-
ments that allowed them to construct the neuronal network of C. elegans:
1. Learned aversive olfactory behavior requires the presence of the AWB and
AWC olfactory neurons. The network shows one candidate network that
regulates this naïve and learned behavior.
134 Intermediate Level Neuronal Networks
2. Two different neuronal networks regulate the attractive (naïve) and aversive
behavior (learned), respectively.
3. The destruction of the AWB sensory neuron in one of these networks abol-
ished the naïve attractiveness and learning.
4. The destruction of RIA interneuron in the other network abolished the
learned preference and learning but did not affect the naïve attractiveness.
(Abstracted and paraphrased from Ha et al. 2010)
about 135,000 neurons compared with these other model animals. This number
currently exceeds the threshold for the kind of neuron-by-neuron analysis that
was possible in the other animals. Current efforts to map out the nervous system
are limited to a few “compartments” and only 16,000 of the 135,000 neurons
of the fruit fly’s nervous system have been identified (Kohl and Jefferis, 2010;
Chiang et al., 2011). To accomplish even this “coarse” and partial wiring diagram
required that Chiang’s group dissect more than a million fly brains! Clearly, this
type of complete neuronal wiring diagram is probably not going to be possible
for an organism as complicated as that of D. melanogaster, much less that of the
human brain in the foreseeable future. Furthermore, even if we had a complete
wiring diagram, the technical difficulties involved in repeatedly stimulating and
recording from a specified neuron (or group of neurons) in the fly are formi-
dable. The neuronal network of D. melanogaster clearly establishes a high limit
for what we have designated as intermediate levels of complexity.
It is also important to remember that in the long history of using D. mela-
nogaster as a means of studying genetic effects on anatomy and physiology,
cognitive neuroscientists are doing something quite different when comparing
behavior and genetic information. Olsen and Wilson (2008) clearly express this
caveat in the following terms:
In other words, behavior is not nowadays being used as a tool for studying
chromosomal effects; rather the goal is to explain the neural foundations of overt
behavior. There is an important difference between studying the genotype and
the neural or behavioral phenotype; the success that has been forthcoming in
the genetic studies sometimes leads us to overstate our accomplishments in the
neurophysiological domain. Furthermore, many of the properties that were
advantageous in the other model preparations (e.g., small number of neurons,
simple coding schemes, isolatable and identifiable neurons, or simplifying con-
siderations like the regularity of the ommatidia anatomy in L. polyphemus) no
longer apply.
Despite these disadvantages, some impressive neuroscientific work relevant to
behavior has been done using D. melanogaster. The fruit fly has some advantages;
for example, it has a wide variety of easily observed behaviors—eating, navigat-
ing, reproduction, etc. A major advantage (countering the increased number of
neurons and anatomical uncertainties about its brain) is that the fruit fly has a
much wider array of behaviors for which neurophysiological correlates can be
sought than is available in some simpler animals. The price one pays for this
136 Intermediate Level Neuronal Networks
expanded behavioral array is that the fly’s nervous system is also more compli-
cated and the difficulty in determining which neurons mediate which behavior
has proven to be a serious constraint on the use of D. melanogaster as a model
preparation in neural and behavioral comparisons.
The initial task, therefore, in planning an experiment to study a neural behav-
ioral relationship is to determine which neurons of D. melanogaster’s 135,000
neurons participate in a given behavior. Olsen and Wilson (2008) proposed three
general methods to link particular neurons to particular networks:
neurons may be involved in controlling this behavior), it has been possible to make
some associations between groups of neurons and specific components of the song.
The five types of neurons and the role each plays in the courtship song are discussed
by von Philipsborn et al. and I briefly summarize the findings here.
Two of the five classes of neurons (P1 and plP10) are not involved directly
in motor control of the singing behavior; instead, they seem to control the deci-
sion whether or not to sing. Neuron class P1, von Philipsborn and her colleagues
asserted, was the key to triggering the pulse song—an analog of “decision mak-
ing.” It possessed a rather short axon and was limited anatomically to the region
of the brain. Similarly, plP10 neurons were also involved in the decision to sing.
The suggestion was that P1 was an input to plP10, conveying information from
a distant ganglion since it (plP10) had a very long axon that terminated in a
distant region of the brain.
The three other neuron classes that were involved in the fruit fly’s courtship
song—dPR1, vPR6, and vMS11—coexist in the thoracic ganglion of D. melano-
gaster and may, according to von Philipsborn and her colleagues, collectively
encode the program controlling the specific properties of the song. Figure 4.5
is a theoretical diagram proposed by them for the connectivity pattern of the
salient neurons in a network of neurons for the control of courtship singing.
The array of challenges involved in carrying out an experiment such as this is
impressive.
It is also important to appreciate that the theory shown in Figure 4.5 is of a
much more conjectural nature than are the three models proposed in the earlier
sections of this chapter.
FIGURE 4.5 The neural network mediating singing behavior in Drosophila melano-
gaster. Five classes of neurons (not individual neurons) are involved. P1 and plP10
neurons control the decision to sing and dPR1, vPR6, and vMS11 control the tem-
poral characteristics of the song.
Reproduced from von Philipsborn et al. (2011) with permission.
A prevailing assumption guiding this work is that these model systems are
similar enough behaviorally to the behavior of higher-level organisms to at least
justify drawing functional inferential analogies between the model and the mod-
eled and, at best, assert that we have found the actual neuronal mechanism of
what are common behaviors.
This line of research, however, is constrained by the fact that although the
observable behaviors may seem to be the same, they may be actually instantiated
by very different mechanisms in the model and modeled organisms. For all
practical purposes these are bottom-up approaches to the role of networks and
this type of research using conventional neurophysiological techniques fails for
networks consisting of more than a few neurons.
Throughout this discussion, the limits of this direct neurophysiological
approach have been apparent because the equivalence of the mechanisms under-
lying the model and the modeled behavior, respectively, remains uncertain.
Although philosophers may blanch at the idea (referring to the logical fallacy
of assuming something is true because it is possible), neurophysiologists will
appreciate that there are innumerable, quite distinct mechanisms that could
produce the same behavior. This harks back to the problem of underdetermina-
tion of the underlying mechanisms if we can observe only the behavior.
Among many other conceptual problems with this direct approach to study-
ing neuronal networks, it appears that many investigators who used the model
preparation approach seem to assume that not only do behavioral analogies
exist between higher and lower animals, but also mechanistic identities in the
form of neurophysiological homologies. Unfortunately, even the most robust
Intermediate Level Neuronal Networks 139
data says nothing about the comparable neuronal network in the higher, mod-
eled animal. That there are alternative possible and plausible neuronal network
mechanisms that could produce responses in different animals based on totally
different neuronal networks seems likely; indeed, our intuition tells us that
there must be such alternative explanations. The advantage of this bottom-up
approach is that it is anchored to neurobiological reality—the mechanisms
described to explain behavior are based on real, tangible, and observable
mechanisms.
An alternative and sometimes useful strategy to understand relatively simple
intermediate neural networks is to work from the top down—to infer the possible
nature of the neural network from an organism’s behavior. Unfortunately, most
such theories are functional inventions comparable to those operating at the psy-
chological level. Models such as these are exemplified by the “biologically inspired
cognitive architectures” (Goertzel, Lian, Arel, de Garis, and Chen, 2010, p. 30).
The important thing to remember about such top-down contributions is that
they are not initially based on neural or neuronally based empirical data; they
are functional mathematical descriptions in which the starting point is behavior
and the end points are plausible mathematical and neural constructs—in other
words, hypothetical constructs. The mathematical representation is invented to
subsequently complete the description by showing a plausible manifestation of
the behavior in neural terms—hypothetical neural mechanisms that could imple-
ment the functions described by the mathematics. These theories are not neu-
rophysiological evidence-based but are top-down theoretical statements based on
behavioral observations rather than neurophysiological ones.
A final and very important finding from the direct neuronal approach is that
none of the methods or approaches discussed here was really successful at unrav-
eling the details of a network. Limulus polyphemus permitted us to examine
results distributed over a few receptors but that had to be simplified by the
regular, almost crystalline, structure of the animal’s eye. In the other experiments,
it was not the network interaction between neurons that was studied but rather
the role of specific and individually identified neurons. The “networks” studied
were, at best, interactions between a few neurons. Interactions were just too
complicated for our instruments, our analytical techniques, and our basic assump-
tions to tease out the details of neuron networks whose number exceeded the
fingers on our hand. In short, only trivial “model” networks could be examined.
Multiple electrode arrays could not reach our goal, nor would limiting ourselves
to what were supposedly simplified (to the point of triviality) model prepara-
tions. Nor, for that matter would the construction of hypothetical entities inferred
from top-down methodologies.
In conclusion, the use of model preparations with simple nervous systems has
allowed us to unravel systems with a few neurons. However, this direct neuro-
physiological approach is severely limited to networks made up of very few
neurons and cannot be expected to solve problems involving realistic networks.
140 Intermediate Level Neuronal Networks
Notes
1. It is likely that the phenomenon had been known years before. Edge enhancement
occurs in the work of artists who antedated Mach by many years.
2. By symbolic, I am suggesting nothing more than that there might be neuronal mecha-
nisms that are much more complex than the simple lateral inhibitory interactions
accounting for the phenomenon in the human brain. We cannot exclude alternative
explanations with the kind of experiments we are able to carry out in L. polyphemus—
nor can we for any of the other model preparations discussed later in this chapter.
5
LARGE-SCALE COMPUTER
SIMULATIONS (THEORIES)
OF COGNITION
5.1 Introduction
In the previous chapter, I discussed an approach to studying neuronal networks
that emphasized the examination of real, but small, neuronal networks using
direct neurophysiological methods. The history of this bottom-up method makes
one point very clearly: the complexities of even a very small network (consisting
of from one to perhaps a dozen neurons) were such that any hope of under-
standing how a network worked with a conventional neurophysiological approach
was not likely to be fulfilled with the tools that were available. It was only special
cases (such as the Hartline Lab’s work on Limulus polyphemus) in which certain
simplifying constraints (repetitive regularity) could be invoked that permitted
larger numbers of neurons to be considered. Even then, traditional neurophysi-
ological methods do not seem promising—there are just too many neurons for
standard analytical techniques to work.
Currently, however, new tools are becoming available that promise to change
not only our experimental methodology but the concepts and assumptions that
are certain to guide theory in particular in the years to come. Some of the most
influential are certain to be modern versions of the digital computer. In principle
and a priori, the tool “par excellence” for studying microneuronal networks is
the digital computer, but not just the simpler machines of the class sitting on our
desks. They may be useful for studying the intermediate level of neural networks
of the kind described in Chapter 4. Instead, computers are needed that have the
ability to manipulate huge numbers of computational elements, each one of which
is capable of representing the activity of a single discrete element—an element
such as a neuron—and to store astronomical amounts of data.
142 Computer Simulations of Cognition
Serial Computers
The modern electronic digital computer represents one of the most influential
technological developments in human history. Its importance ranks with the
bow and arrow, animal domestication, the wheel, the microscope, and the auto-
mobile as vehicles for not only technological but also social change. Increasingly,
in modern times, computers have affected the way we live and promise to have
even a greater impact in the future. Especially in the past half-century, computer
power seems to be evolving at a rate that could not have been imagined only
a few years ago. Computer capability has grown so much that it now possible
to dream, if not to achieve an actual simulation of human cognitive processes
at the microneuronal level—thus activating a sea-change in cognitive neurosci-
ence theory development. This is not being done at a symbolic or indirect level
of analysis at which psychology and macroscopic theories operate, but at what
many agree is the level of operation at which the brain actually accomplishes
its wonderful, albeit mysterious, conversion of the neurophysiological to the
mental. To understand the modern supercomputer fully, it is useful to give a
brief history of some of the key points in its development.
Although computing engine-like devices of one kind or another have been
used for many years (e.g., the Roman and Asian digital abacus, Napier’s Bones,
or the 17th century slide rule—an analog device) and a few mechanical calculat-
ing engines (e.g., the Jacquard loom) had been invented over the years, the
modern electronic digital computer that will dominate the following discussion
was first conceptualized and built by Atanasoff and Berry at Iowa State University—
a usually overlooked history told by Burks and Burks (1989). The Atanasoff-Berry
machine was an all-electronic, special-purpose, binary device capable of solving
linear equations by numerical methods—the first of its kind. However, it had one
critical missing feature: It could not be rapidly reprogrammed as it was hardwired
to perform only a single function.
144 Computer Simulations of Cognition
CPU
Arithmetic-Logical Unit
INPUT OUTPUT
Sequence Controller
MEMORY
FIGURE 5.1 The organization of the single-core von Neumann computer composed
of input device, output devices, memory, a central processing unit (CPU) that consists
of an arithmetical-logical unit, and a program sequence step controller. Parallel
processing computers may have many of each of these units connected in a variety
of ways.
Computer Simulations of Cognition 145
The block labeled INPUT consisted of any number of different devices that
are able to take the information stored on some physical medium and enter them
into the memory of the computer. Punched cards, punched paper tapes, teletypes,
keyboards, cameras, and, nowadays, such devices as high-capacity “thumb” memories
or highly sensitive cameras have been used as INPUT devices. Output devices
such as mechanical printers or cathode ray tubes were among the early means of
communicating the results of computations to the outside world. Nowadays such
exotica as three-dimensional printers have begun to be used as output devices.
The central processing unit (CPU) was the computational device that accessed
input data, manipulated it, and prepared it for presentation to the output devices.
The CPU is often depicted as composed of two parts—an arithmetical-logical
unit that actually executes the stored instructions and a control unit that tells
the computer what instructions should occur next as it sequences the various
steps of a program. Another name for the CPU was the “core,” a term that
rose in popularity as the older notion of a simple, single CPU began to be
replaced by parallel processing systems that had multiple CPUs or specialized
data manipulation devices that were quite different than the conventional single
CPU of the classical von Neumann computer.
Memory technology evolved over the years from a novel use of Crookes
tubes, to small magnetic spots on a drum, to vacuum tubes, to discrete solid-state
circuits, to a huge variety of highly miniaturized and integrated electronic circuits,
capable of rapid reading and writing of vast amounts of stored information. The
development of fast printed memory circuits that could stably preserve informa-
tion with or without power has been one of the grand successes of computer
memory technology. Where originally a couple of thousand volatile memory
locations constrained the size of programs, modern memories consist of gigabits
of stable and rapidly accessible memory locations.
Early computers based on vacuum tube technology or early solid-state circuitry
were ponderously slow by today’s standards. They were all serial machines in
which each instruction or computer access had to be completed before the next
one could be executed. The main practical strategy to speeding up the execution
of a computer program in the early years was to improve the speed at which
memories could be accessed and information processed by the CPU. This led
to an emphasis on engineering development of faster and faster technology (i.e.,
the electronic components); however, the basic organization plan of the von
Neumann machine remained more or less constant, in general, if not in detail.
Eventually, physical limits of heat generation and disposal and the basic physical
limits on electrical conduction time (i.e., speed of light considerations) placed
limits on how fast circuits in this kind of computer could function.
Refrigeration helped in part to overcome heat problems and new semicon-
ductor materials helped to speed up computers, but eventually the physical limits
encountered with very small circuits seemed to place limits on how fast a serial,
single CPU computer could operate. Nevertheless, the first few generations of
146 Computer Simulations of Cognition
computers were all serial computers with a single CPU that was driven to the
physical limits of available circuit design speeds. This did not mean serial com-
puters were paralyzingly slow for their times; very large systems (e.g., the Cray-1
of 1976 with only a single CPU-dominated computer system speed until the
1970s) could accomplish much by concentrating on speeding up a single CPU.
We can sum up the extraordinary properties of digital computers by listing some
of the most significant ideas that emerged during the past half-century that influ-
enced the history, design, and applications of modern computer engineering:
The physical limits on how fast a serial computer could process information
were encountered just as the real world was demanding solutions to problems
that required much faster computing speeds. Weather forecasting, molecular
interactions, and a host of other civilian and military applications arose that could
not be satisfied by the fastest available conventional single-CPU computer. The
engineering solution that finessed the physical limit barrier was based on the
concept of parallel processing—an idea that leapt into prominence in the 1980s.
This is the topic of the next section.
the two. Special computational units in the earliest supercomputers were added
to what were otherwise serial computers in an effort to speed up computational
throughput. Later attempts to build ultrafast supercomputers made a dramatic
change when it became possible to “print” large numbers of CPUs (the number
now approaching millions of these no longer “central” but now better described
as “distributed and parallel” processing system processing units). In recent years
it has also been possible to build virtual supercomputers (e.g., the Beowulf system
described in Becker, 1995) by taking advantage of world-wide communication
links to employ distributed networks of ordinary desk computers to collectively
solve the kind of problem that could be partitioned into separate tasks. The
fastest supercomputers, however, still strive to package many memory and pro-
cessing units in a single location; otherwise, communication times can become
troublesome. The best solution to this difficulty is to concentrate on problems
in which the partitioned components are totally independent. Unfortunately,
nature does not often serve up its problems so neatly.
During the past decades of the 20th century, the idea of designing parallel
processing computers with very large numbers of multiple cores and huge arrays
of memory storage units became increasingly popular, and substantial investments
were made by corporations and governments to engineer ultrafast computers.
Parallel processing computers have evolved into a wide variety of system types;
most fundamental differences are involved in the manner in which the multiple
cores are interconnected with one another and with the memory storage units.
Cores can be interconnected so that they are clustered together but still operate
quasi-independently. Parallel processing computers also may have a variety of
different means of accessing the memory. In the simplest cases, in which the
memory is shared by the two CPUs, both might have full access to all of the
memory, while in other designs only limited access to the common memory is
available to each core. In a system in which two or more memories are con-
nected separately to their respective CPUs, a common communication “bus” is
necessary to let the two memories connect to each other. From a primitive
two-processor, shared memory, the parallel processor has evolved a substantial
variety of computer architectures in which multiple memories and CPUs are
interconnected in complex ways by ever more complicated buses. Although it
is impossible to provide a complete list of all possible parallel processing archi-
tectures, a simple and inclusive taxonomy was presented by Flynn (1972) that
characterizes the main types. He categorized the main types but acknowledged
that there were many variations on these main architectural types:
These idealized examples do not tell the full story because some tasks are only
partially parallelizable. That is, in most problems, only a portion of the program
can be broken down in a way that would permit additional CPUs to speed up
the solution of the problem by bringing more raw computer power to the task.
Indeed, there is a classical relationship known as Amdahl’s (1967) law between
the maximum possible speed increase and the proportion of the program that is
parallelizable. A much simplified version of Amdhal’s law asserts that the maximum
possible speedup is less than 1/1-P for a given number of processors, where P is
the proportion of the problem that can be parallelized. Thus, the maximum pos-
sible speedup of a task is not only a function of how many processors might be
present in the computer (which in a perfectly parallel problem would be a linear
function of the number of available cores), but also of the nature of the posed
problem. As an example, if the problem was determined to be only about 50%
parallelizable, then this law would predict that the fastest parallel computer would
only be able to speed up program execution by a maximum factor of 2 over that
achievable with a single ultrafast CPU computer. This would hold no matter
how many CPUs were assigned to work on the problem.
In addition, the execution of a parallel program typically involves a number
of additional housekeeping steps that were not required in a serial computer
(e.g., assigning parts of the problem to particular subsets of the hardware and
communicating data to multiple locations). In sum, one cannot expect to reach
the theoretical maximum improvement suggested by Amdahl’s law; there will
also be some housekeeping costs to pay.
Nevertheless, with the extraordinary number of highly miniaturized cores
that are now being combined into today’s supercomputers, an enormous amount
of computing power is becoming available that could not be achieved with
simply technological (i.e., individual circuit speedup) changes. Some of the most
powerful systems claim that they are composed of multiple millions of cores
simulating tens of thousands of neurons. In addition to the economic, military,
and scientific implications of high-speed computing, there are now quasi-political
overtones and international competitions. Over the past few decades there has
been an international race to build the “fastest” computer and for each competi-
tor the strategy of choice has been to instantiate the contending computer in
the form of a massively parallel system and then to test its performance with a
standardized program, an example of which is Linpack, a program designed to
solve a “dense system of linear equations.”1
The metric of computer speed used to compare the contenders in this race
is the unit referred to as Floating Point Operations per Second (FLOPS). As a
raw measure of computer speed, the number of FLOPS is preferred to the number
of machine language program instructions carried out per second because it
represents a more realistic estimate of problem-solving ability, especially in float-
ing point’s ability to represent very small and very large numbers and fractional
values (integer arithmetic deals only with whole numbers).
Computer Simulations of Cognition 151
Name FLOPS
YettaFLOPS 1024
ZetaFLOPS 1021
ExaFLOPS 1018
PetaFLOPS 1015
TeraFLOPS 1012
GigaFLOPS 109
MegaFLOPS 106
KiloFLOPS 103
the largest parallel computer still does not come close to the 86 billion neurons
of the human brain, it has been possible to use a smaller number to simulate
numbers approaching that of the brain by means of iterative programming.
Therefore, there is at least a superficial comparability that suggests that the basic
functionality of the brain might scale down to that of the best modern comput-
ers or the computer might scale up to the level of complexity of the human
brain. The current world speed holder—the Chinese Tianhe-2 supercomputer—
consists of approximately 3 million separate cores of several different types. IBM’s
Sequoia supercomputer, the record holder in 2012, was built from 1.5 million
cores.
It is important to distinguish among the number of electronic neurons, cores,
and electronic circuits in a supercomputer simulation of a neuronal network. At
the bottom of the hierarchy is the number of electronic circuits or gates. The
number of these electronic circuits can be very large. Modern printed circuity
has reduced the size of a binary flip-flop (the basic logical circuit) to molecular
levels. Millions of these circuits are present in the most advanced supercomput-
ers. On the other hand, a much smaller number of cores or CPUs are constructed
from these basic circuits. Several million equivalents of cores seem to be the
frontier currently. In turn, a number of cores must work together to represent
a neuron so that a neuronal network can be simulated. About a million synthetic
neurons is the best so far achieved. What is clear is that the number of actual
physical neurons in the synthetic network is still significantly smaller than in the
real biological brain and thus, repeated cycles of program execution would be
required for the computer to run as fast as the brain.
Thus, there still remains a scale difference between the number of neurons
in the brain and the number of cores in the most powerful supercomputers. It
is not certain whether this difference will be significant or whether computers
can be scaled up to the count of biological neurons. The important thing is
that the past decade is the first period in cognitive neuroscience history that the
basic organization properties of the technological tools available to study the
brain and those of the brain itself are commensurable. The von Neumann serial
computer under control of a programmable sequencer and a single computational
element is fundamentally different in this regard from how the parallel processing
computer and the brain operate. It was a poor model on which to base theory
development, not only because of its serial nature, but also because of the dif-
ference in how the fundamental serial concept is embodied in the hardware.
Although the degree of parallelicity is still far greater in the brain than the best
computer, at least the basic principles of operation are now apparently converg-
ing. The basic concept defining the organization of a parallel processing computer
and of the brain match much better now than with any previous technology.
Computers and the brain share another basic similarity that makes the parallel
processing computer nearly the ideal tool for studying the brain. Unlike the serial
computer in which each instruction is controlled in sequence by the commands
154 Computer Simulations of Cognition
the basic concept that permits the brain to carry out operations that would have
been impossible with the intrinsic slow speed of the neuron. Single cell theories
of cognition are, from this point of view, simply incorrect at the most basic
conceptual level. It is not possible for any single neuron to carry out the
information-processing speed represented by our amazing behavioral abilities
without massive parallelicity. The hope is that the new technology will open
the door to future understanding, perhaps even crack the mind–brain problem.
However, there are formidable obstacles to the construction of an overarching
theory of how the brain produces the mind.
One such obstacle is that even the most advanced supercomputers are cur-
rently limited in the number of individual CPUs that have been joined together;
the 2 or 3 million that have so far been combined are still far short of the
number of neurons in the brain. Complexity increases with numerousness,
therefore, there is no assurance that the emergent properties we designate as
cognitive will appear with the supercomputers that are available at the moment.
It is likely, if not certain, that we may still not be at the appropriate level of
analysis and may still have to depend on future supercomputers with many bil-
lions of synthetic neurons before we can claim a solution.
Another obstacle is the problem of interpreting the results—when will we
know that we have built the desired overarching theory of mind–brain equiva-
lence? The output of any system, brain or computer, is presented in some form
of behavior—for example, a successful pattern recognition. It is possible, however,
to produce behaviors by several different means (where the word “several” may
indicate very large numbers; see the discussion of Hilgetag, O’Neill, and Young,
1996). We call these behaviors that are produced by different mechanisms and
processes “analogs” of each other. The problem is that the externally observed
behavior of a person or a computer is underdetermined; there are not enough
data in the behavior to determine which of “several” possible mechanisms
accounted for that behavior. Both computers and people can be directed to
produce particular behaviors in innumerable ways.
How then do we determine that a particular underlying mechanism (a reduc-
tive theory) is actually explaining the behavior in humans and not just simulating?
The answer to this rhetorical question is that there is no way of choosing among
the analogs. However accurate a theory may be in producing a particular behavior,
it may do so by mechanisms that are not only completely different than those
used by the brain but may also be inaccessible to the most powerful of our tools.
Recognizing a face, for example, can be accomplished by a computer by measur-
ing the connecting lines between a set of key points (e.g., the tip of the nose)
and then creating a numerical vector associated with that particular face. Humans
also recognize faces but it seems extremely unlikely that we do so by comparable
numerical calculations and measurements. Rather, cryptic (so far), organizational
processes involving molar spatial relationships of the kind studied by the
nearly forgotten Gestalt psychologists seem more likely mechanisms of
156 Computer Simulations of Cognition
As the end of the 20th century approached and passed, this situation began
to rapidly change. Massively parallel supercomputers became available, and sci-
entists and engineers of all persuasions quickly took advantage of their presence
to execute large-scale simulations.
As promising as the parallel processing supercomputer is, in principle, as an
engine of theory building, it is also possible to foresee some of the difficulties
that are being faced and will continue to be faced in their application to the
task of building cognitive neuroscience theories. Some of these difficulties are
subtle and some are naïve confusions between the search for an explanatory
theory, on the one hand, and the engineering of bigger, faster computers, on
the other. Merely noting that a computer engineering project is “inspired by”
a brain property does not a mind–brain theory make.
An initial problem facing theoreticians who use massively parallel processing
computers as their tool of choice to theorize about cognitive or brain activity
is that it is difficult to know what would constitute an acceptable explanatory
theory. There are two source of this difficulty. First is the fact that both the
overall internal states of a computer and mental activity are largely inaccessible.
We have only the most indirect indications—mainly interpersonally observable
behavior or introspective reports of the cognitive concomitants of a brain activity.
However, neither of these measures permits access to the lower-level mechanisms
that constitute the “meat and potatoes” of any reductive theory.
Thus, the most abstruse entity of all—consciousness—remains poorly defined
and immeasurable as well as inaccessible; neither overt behavior nor covert
thought provide any direct indication of what are the underlying mechanisms,
either neurophysiological or hypothetical cognitive constructs that correspond
to our awareness. As noted repeatedly in the course of my writing, behavior
is underdetermined—it can be produced by what may be an innumerable
number of possible internal mechanisms and there is insufficient information
in a behavioral observation to determine the actual nature of those mechanisms.
Therefore, we may arrive at a point where we have a sufficient answer to the
basic query but no way of providing the necessary answer. If plausible sufficiency
is acceptable to some theoreticians as the criterion for acceptance of a (or the)
theoretical explanation, then so be it. It must be remembered, however, that
plausible sufficiency is not too distant from the pejorative term “folk psychologi-
cal theory”—it is just one of a possible multitude of mechanisms that can
produce analogous behavior. One implication of such an interpretation is that
if some kind of “necessity” or “truth” is desired, our search for a theory may
be a long one.
Another major difficulty in basing theories on computers is that the very
processes in which we are interested may emerge at a level of system complexity
that exceeds even the most powerful computer’s ability to manipulate the relevant
information. That is, whatever the cognitive process under study, it simply may
not be present at the levels of analysis at which we are capable of problem
Computer Simulations of Cognition 159
solving. To make this subtle point more clearly, whatever consciousness is, it may
not exist in a network (neural or otherwise) until a sufficiently high level of
network complexity exists and then, for computational reasons, it may be impos-
sible to either produce or evaluate an overarching theory of the mind. Only
when network complexity rises to an unknown, but extreme level, may a cogni-
tive process actually become manifest. One has only to compute the factorial
of a relatively modest number to appreciate that technological and combinatorial
inadequacy in the face of realistic complexity may persist as a problem for
cognitive neuroscience into the foreseeable future. (For a more complete discus-
sion of this issue, I refer my readers to Chapter 3.)
I am most interested in the critique of the strategy of trying to draw overall
conclusions of the individual’s nature from studying the social algorithms.
The point of this list in the present context is that not all of these expected
outcomes may be available to researchers for both practical and conceptual
reasons. Nevertheless, it is at this level (that of the individual states of the under-
lying individual neurons and synapses) that the essence of our cognitive activity
is most likely to reside. It is not clear, for example, how one could actually carry
out an experiment in which the experimenter had sufficient control over the
independent variables to erase or produce a particular pattern. Experiments
calling for “induction” or “saturation” sound very much as if they were mac-
roneural manipulations rather than the microneuronal network manipulations
that characterize parallel processing networks of the kind with which we are
concerned in this chapter.
There is another very important question that has to be considered about
each parallel processing theory—what are the respective goals of the various
supercomputer “brain” models considered here? Although this book is aimed at
considering those that specifically link to cognitive and neurophysiological stud-
ies, this is not necessarily the basic motivation of all of the models discussed
here. Some, as we shall see, are primarily interested in large-scale simulations of
neuronal networks, some with the production of particular useful behaviors, and
some in the engineering of massively parallel computer systems themselves. The
Computer Simulations of Cognition 161
cerebral cortex. Others have had as their goal simulation of the interactions
that may occur between other anatomical structures such as the thalamus and
the cortex. So far, however, in this listing, explaining cognitive processes has
been almost completely absent.
Another class of parallel processing computer research, however, is aimed spe-
cifically at comparisons between computer architecture and cognitive processes.
Nevertheless, in this class of research, there is less interest in neurobiological fidelity
than in producing comparable behavioral outcomes. In this context, the goal is
to make computers behave like people, but not to restrict or constrain programs
or hardware algorithms to the same computational processes used in the brain.
Needless to say, this imitation, simulation, or emulation (and often enhancement)
of human behavioral processes has been the main thrust of modern applications—to
supplement or replace many of the tasks that previously had been carried out
exclusively by humans without necessarily reproducing the same logical
“programs.”
As one considers this spectrum of motivations and despite the enormous
interest in linking brain anatomy and neurophysiology to cognitive processes, it
is important to appreciate that no cognitive neuroscience problem of mind and
brain relationship has been resolved so far by the expensive and complex research
that has been forthcoming; most theories of computer organization that have
any relevance to cognition in computer science are still “toy problems.” Clearly,
there is a multitude of ways that computers can simulate human behavior that
do not require the construction of exactly the same computer operations and
procedures that the brain presumably uses. Simulations and imitations, no matter
how perfect, cannot by themselves provide a necessary explanation of how the
brain works to produce cognitive activity—merely a sufficient imitation of
behavior. To the contrary, the ultimate theoretical goal of cognitive neurosciences
is to understand the brain well enough to mimic behavior to a degree of detail
that permits us to assert that we have solved the mind–brain problem.
This leads to another strategy in modeling cognitive processes with large-scale
supercomputers—what we may refer to as the “hopeful” approach. The technique
embodied in the hopeful strategy is to develop a simulation of a prototypical
neuron and then connect large numbers of them together and just see what
happens. It is probably the case that more of the parallel processing prototheories
currently being considered fit into this category than is admitted by investigators
in this field. Why this should be the case is obvious—although we know a lot
about neurons and synapses, we know very little of the rules governing their
transition to behavior. On the positive side of this “hopeful” scenario is the fact
that an enormous amount has been learned about the engineering organization
and performance of large-scale, parallel processing supercomputers.
There can be no dispute, however, that the most likely tool currently available
to achieve any of these goals is likely to be some future descendent of the parallel
processing supercomputer we know today. In this cognitive neuroscience context,
Computer Simulations of Cognition 163
both the individual cellular and “macrocircuit” levels that determined the kinds
of oscillatory behavior resulting from these interventions.
Lumer, Edelman, and Tononi’s report is especially interesting because it is
unusually complete in spelling out the details of their goals, methods, results, and
conclusions. It is clear how they go about defining the initial conditions—the
stimuli—and the cumulative or population-averaged response measure they used
to examine a very specific property of networks—its oscillatory behavior. From
the neuroscience point of view it is anchored to well-established neurophysiologi-
cal data concerning the structure and function of the early portions of the visual
system. Thus, it added a significant amount to our understanding of how neural
networks might be affected by measurable properties of the nervous system. The
anatomical properties of the model correspond to well-known properties of the
visual system such as layering and feedback.
On the other hand, their model is limited, as they themselves point out, in
that it does not link directly to many of the problems of perception—instead it
is a test bed to study biologically plausible networks of a well-defined portion
of the nervous system. But they have not correlated these plausible visual neu-
rophysiological properties with psychophysical data. In their words:
While the simulations presented in this paper do not yet address func-
tional questions . . . they provide a powerful tool to begin investigating
these [perceptual] phenomena in closer contact with experimental
observations.
(Lumer, Edelman, and Tononi, 1997a, p. 224, italics added)
Spiking Neurons
Another early project in which the stage was set for the more recent and much
larger computer simulations was reported by Izhikevich (2003). Izhikevich
programmed a relatively simple network model in which each simulated neuron
(of a set variously mentioned as including 1,000, 10, 000, and “tens of thousands
of simulations”) was represented by a set of simple differential equations. The
simulated neurons were interconnected by up to a million simulated synaptic
connections, 80% of which were excitatory and 20% of which were inhibitory.
166 Computer Simulations of Cognition
Over the past decade, the IBM part of the project has gone in a different
direction—they have aimed at developing novel computer architectures with
ever higher numbers of cores with faster and faster processing capability. Using
the Sequoia supercomputer, one of the most powerful parallel processing super-
computers available in 2012, the IBM group was able to execute a program that
they claim simulated a system of 500 billion neurons and 100 trillion synapses
(as described in the article by Service, 2014). This had to be done with an itera-
tive process since the number of actual cores was only 1.5 million and the logic
was built on relatively slow conventional integrated circuits. As a result, the
program ran much more slowly than did estimates of the parallel processing
speed of the human brain. The Sequoia project was not designed to simulate
the human brain; it was mainly intended to develop an efficient computer
hardware system that was “inspired” by the brain.
The next step in the development of such a system was to engineer a unit-
ized “neuromorphic” large-scale integrated circuit chip that incorporated the
processing features of many neurons by means of specialized processing cores.
The IBM TrueNorth device that emerged from these primarily engineering
goals was a single chip that contains 5.4 billion transistors and 256 million simu-
lated synapses simulating the behavior of a million simulated neurons. The logic
design was organized in a manner that was assumed to be comparable to the
complexities of the human brain. Current research reports suggest that it has
been possible to link “more than a dozen” of these integrated circuit chips
together, but the developers feel that this is only the beginning and that the
system is capable of unlimited upwards scalability by simply attaching more and
more of the TrueNorth chips to one another.
One of the most interesting aspects of the TrueNorth project is that the cores
on these chips are not based on the serial von Neumann prototype. That clas-
sical design placed the core and the memory in different locations, resulting in
high-energy demand whenever information was transmitted between the memory
and the cores. This limitation of the von Neumann–type core could be abetted
by placing memory and cores on the same chip, thus reducing transmission times
and resulting in a remarkably low power demand. The IBM group (Merolla
et al., 2014) reports that the TrueNorth chip drew only 63 milliwatts.
The development of the TrueNorth chip was both an important milestone in
packing circuits onto a single chip and novel logic design. However, it consisted
of only a relatively small number of simulated cores and neurons in comparison
to the 2012 Sequoia-based project. What it accomplished should not be confused
with the kind of progress made in the 2012 project. The 2012 project carried out
programs on a supercomputer that simulated numbers that actually exceeded the
biological brain’s neuron counts (530 billion) but had to do so by repetitive use
of the circuitry in the Sequoia computer they utilized and thus not in real time.
The ultimate goal of the TrueNorth project, on the other hand, is to design
circuitry that mimics these large numbers of neurons in the brain and pack it
172 Computer Simulations of Cognition
onto a single chip—a chip that could be scaled upwards by interconnection with
other TrueNorth chips to represent billions of neurons. The TrueNorth chip has
properties that may ultimately achieve this goal, but it has not yet been expanded
to approximate the number of neurons in the brain—86 billion.
The IBM TrueNorth project, although motivated mainly by engineering goals,
also has been used to simulate some relatively simple cognitive processes. Inputs were
exemplars of a set of moving objects. The behavioral goal of the system was correct
detection and recognition by the TrueNorth chip of these different objects; however,
no performance data were provided in the brief articles describing the TrueNorth-
based simulations nor were the methods used to input the initial conditions or to
develop program modification (i.e., learning) as a result of the activity in the chip.
In summary, the HBP and the TrueNorth projects had many similarities in
their goals; however, they also were quite different in some ways. Specifically,
the main goal of the HBP is to program whatever computer system is available
to simulate billions of neurons in the human brain and to do this at multiple
levels of analysis. As each level is simulated, the intent is to concatenate these
various simulations until a neural “model” of the brain emerges. The target in
this case is the neuroanatomy and neurophysiology of the brain itself.
The IBM work is, on the other hand and quite to the contrary, aimed at the
development of hardware, inspired to a degree by the biology of the brain, but
more concerned with such functional properties as power efficiency, fault toler-
ance, packing density, and scalability, as well as programmability. The brain, to
the IBM group, is only a distant benchmark to be implemented someday by
complex circuitry—the simple behavioral functions are merely a means of dem-
onstrating the capabilities of the system. Should there be any insights about the
brain and specifically about what artificial cognitive functions are obtained during
this engineering-motivated development project, that would be “icing on the
cake.” In the meantime, allusions to the brain-like features of all of these super-
computers might well be taken with extreme caution.
relations in the published literature. Indeed, we have not yet learned how to carry
out a relevant experiment on these supercomputers. A major problem inhibiting
such accomplishments is the difficulty of programming these computational behe-
moths. In actuality, the task is comparable to the analogous development process
that occurs at least over the lifetime of an individual human and possibly over eons
of organic evolution. Indeed, it is not even sure what the word “programming”
would mean in the context of a computer that operates in parallel on many com-
ponents of the input and responds in the form of a distributed network.
In large part, whatever the actual motivations and goals, many supercomputer
engineering projects are being promoted on the basis of promises and hopes in
the fields of neurophysiology and psychology. Although the power of the brain
is clearly a function of the myriad synaptic connections, we still have no idea
what are the specific patterns of these connections that correspond to behavior
selection, much less consciousness. Without question, newer and faster computer
systems are going to emerge from this activity, but we are still in need of a
conceptual revolution to determine how the parallel processing supercomputer
will contribute to cognitive neuroscience.
Surveying this chapter, we may draw a number of tentative and preliminary
conclusions about the promising role of these new computers as theory engines:
8. We do not yet have a solid idea of how many neurons in the brain are required
to carry out cognitive processes or how many cores are necessary to solve a
computer problem.
9. Large-scale parallel processing computer research is being carried out with
multiple motivations and goals. Most are aimed at engineering development
or simulation of particular neuronal networks. Often, research in this field is
embellished with words such as “brain” or “neuron,” but the cognitive and
neurophysiological meanings of these words is long lost. Many such “brain”
projects have little to do with the goals of cognitive neuroscience.
10. Looming on the horizon is what is probably the most significant practi-
cal obstacle of all—the combinatorial explosion. The enormity of combina-
tions and possibilities is not yet a part of the thinking of most psychologists
and physiologists. There may well be things we cannot do because it would
require computers as large as the knowable universe.
11. Similarly, the criteria for demonstrating equivalence of brain function and
cognitive processes are far more stringent than generally understood. How
will we know when a simulation provides a necessary explanation and not
just a sufficient one?
12. A sampling of the large-scale parallel processing simulations being carried
out in recent years displays the diversity of their goals. Some are strictly engi-
neering, some are neurophysiological, and a few aspire to being models of
cognition. Most end up being examples of cumulative neuronal activity that
are not linked to cognition. Some end up being restricted to sensory pro-
cesses rather than anything that might be called cognitive. Some, after years of
research and huge amounts of money expended, end up with nothing of rel-
evance to cognitive neuroscience. The words used to describe one’s favorite
project often do not discriminate between these disparate goals and findings.
Notes
1. It is important to remember that not all “neural” computers engage ideas from neu-
rophysiology. The mathematics specific to neural-like networks is characterized by a
goal of dealing with networks of linear threshold units. Engineers are concerned with
increasing the speed and power of computers using strategies that may range from
simple metaphors of “brain-like” networks without any specific interest in building a
Computer Simulations of Cognition 175
theory of cognition. In this book, I have striven, not always successfully, to find examples
of projects that have specific cognitive neuroscience implications.
2. About this time, I (Uttal, 1975) proposed an autocorrelation theory of form detection
that was also a parallel processing system. Calculations of the model were successful in
predicting the detectability of dotted forms hidden in dotted noise.
3. The Xeon computer CPU is a modestly powered core used in a system that is just
barely above a simple desk computer. It is not possible to estimate the speed at which
such a computer might run a 2.5 million-neuron simulation. However, it is very clear
that, even with eight cores, it would not be comparable in execution time to any of
the high-level parallel processors discussed elsewhere.
4. Some glimmering of hope for the development of super displays is on the horizon.
Ahrens et al. (2013), for example, have developed a technique combining optogenetic
and optical techniques to image the entire body of a larval zebrafish. This permitted
them to simultaneously show 80% of neurons in this model preparation at a resolution
that allowed them to scrutinize individual neurons. Although promising a considerable
technical accomplishment, this technique does not solve the problem of numerousness
(the data explosion) nor of complexity (the combinatorial explosion).
5. Although the same criticism may be made of the United States–sponsored BRAIN
project, that very large project has much more modest goals—namely to provide the
instruments necessary to “reverse engineer” the brain at some future time.
This page intentionally left blank
6
EMERGING CONCLUSIONS
This book has been aimed at partially answering one practical question and one
theoretical question. The practical question is—how does the available technol-
ogy affect microneuronal theory development in cognitive neuroscience? To
answer the first question, it has been deemed to be appropriate to survey the
nature of the various kinds of instruments that are available, the kinds of empiri-
cal results that have been forthcoming from their application, and the relevant
theories. It is clear that remarkable progress has been made in both instrumenta-
tion and empirical neurophysiological research and that the impact on contem-
porary theory has been substantial.
The main answer to the second question—the theoretical one—that we must
draw from the review carried out in this book is that despite substantial progress
in understanding the brain, we still have no idea of how the mental emerges
from the physiological. What we have, instead, is an unsubstantiated but increas-
ing appreciation that the key to solving the theoretical question is going to be
found at the microneuronal level, that is, at the level at which the properties of
the individual neuron or of networks of neurons are preserved. It is not to be
found, as I argued in my companion book (Uttal, 2016), at the macroneural
level that characterizes so much of current brain-imaging research. It is not, I
argued there, that the cumulative, statistically summarized, pooled macro-responses
that purportedly represent the codes for cognition, but, instead, the nature of
individual neurons and their combinations and interactions.
This book has surveyed the technology of microneuronal research and the
kinds of theories that are driven by observations that are obtained at this micro-
neuronal level. It is a truism that none of these methods is perfect and that
each is flawed in some fundamental manner or is still in an early stage of
development. Nevertheless, much of what we obtained at the level of single
178 Emerging Conclusions
12. The idea of a network of many interacting neurons has begun to dominate
and replace theoretical thinking in cognitive neuroscience. This approach
was stimulated by the pioneering work of Donald Hebb, who postulated
networks of neurons organized into functional groups whose configurations
varied with the state of the involved synaptic junctions.
13. Unfortunately, the Hebb conjecture is probably untestable since it requires
enormous amounts of information about the detailed structure of the net-
work. No conceivable neurophysiological methodology permits experiments
of this kind in which either adequate data are accumulated or stimulus con-
figurations are properly controlled.
14. Some (how many can only be guessed at) cognitive neuroscientists seem not
to appreciate the basic principles of combinatorial arithmetic. A very small
sample of nodes (neurons) can produce astronomical numbers of combi-
nations—numbers that often turn out to exceed the space available in the
known universe or the durations that exceed the known age of the universe.
The possibility of an exhaustive analysis of a system like the brain is remote,
perhaps even demonstrably impossible. Even such superficially simple tasks
such as spike action potential clustering or the determination of functional
connectivity maps may be impossible because of combinatoric constraints.
15. The direct neurophysiological approach in which the arrangement and
states of the neurons involved in a real neuronal network is also unlikely
to be achieved. It is not possible in even the most reduced model prepara-
tion to track out the functions of a network consisting of more than a few
neurons.
16. Furthermore, it is never certain whether a mechanism observed in a model
nervous system is implementing the same function in the same way as the
ones underlying human thought. Indeed, it is more likely in a simplified
model preparation that we are looking at analogs rather than homologs of the
processes of interest.
17. Redundancy is a basic fact of life in the nervous system. An observed neu-
rophysiological mechanism may be only one of many alternative (and con-
stantly shifting) structures.
18. The parallel processing supercomputer is achieving numbers of synthetic or
programmed neurons that are so great that it is feasible to begin to think of
the possibility of simulating human brains. However, loading a supercom-
puter with initial conditions and recovering the myriad details of the resulting
response remain major difficulties. However, this supercomputer approach is
only a few years old and much is to be learned about what they can and can-
not offer to the problems of cognitive neuroscience theory development.
19. Even a successful simulation of a cognitive process on a supercomputer would
not solve the problem of distinguishing between a sufficient and clever ana-
log, on the one hand, and a correct and necessary explanatory explanation
Emerging Conclusions 181
Bakker, M., & Wicherts, J. M. (2011). The (mis)reporting of statistical results in psychol-
ogy journals. Behavior Research Methods, 43, 666–676.
Bandettini, P. (2007). Functional MRI today. International Journal of Psychophysiology, 63(2),
138–145.
Bargmann, C. I., Hartwieg, E., & Horvitz, H. R. (1993). Odorant-selective genes and
neurons mediate olfaction in C. elegans. Cell, 74(3), 515–527.
Barlow, H. B. (1953). Summation and inhibition in the frog’s retina. Journal of Physiology-
London, 119(1), 69–88.
Barlow, H. B. (1961). Possible principles underlying the transformations of sensory mes-
sages. In W. A. Rosenblith (Ed.), Sensory Communication (pp. 217–234). Cambridge,
MA: MIT Press and New York: John Wiley.
Barlow, H. B. (1972). Single units and sensation: A neuron doctrine for perceptual psy-
chology. Perception, 1, 371–394.
Barrett, L. F. (2011). Bridging token identity theory and supervenience theory through
psychological construction. Psychological Inquiry, 22, 115–127.
Barrett, L. F. “What Emotions Are (and Aren’t).” The New York Times, August 2, 2015.
Barrett, L. F., & Satpute, A. (2013). Large-scale brain networks in affective and social
neuroscience: Towards an integrative architecture of the human brain. Current Opinion
in Neurobiology, 23, 361–372.
Bashir, S., Perez, J. M., Horvath, J. C., Pena-Gomez, C., Vernet, M., Capia, A., . . . Pascual-
Leone, A. (2014). Differential effects of motor cortical excitability and plasticity in
young and old individuals: A Transcranial Magnetic Stimulation (TMS) study. Frontiers
in Aging Neuroscience, 6(111), 1–13.
Becker, D. J. (1995). Beowulf: A parallel workstation for scientific computation. Paper presented
at the International Conference for Parallel Processing.
Belliveau, J. W., Kennedy, D. N., Mckinstry, R. C., Buchbinder, B. R., Weisskoff, R. M.,
Cohen, M. S., . . . Rosen, B. R. (1991). Functional mapping of the human visual-
cortex by magnetic-resonance-imaging. Science, 254(5032), 716–719.
Belykh, I., de Lange, E., & Hasler, M. (2005). Synchronization of bursting neurons: What
matters in the network topology. Physical Review Letters, 94(18).
Benjamin, B., Gao, P., McQuinn, E., Choudhary, S., Chandrasekaran, A. R., Bussat, J. M.,
. . . Boahen, K. (2014). Neurogrid: A mixed-analog-digital multichip system for
large-scale neural simulations. Proceedings of the IEEE, 102(5), 699–716.
Bennett, M. R. (2001). History of the Synapse. Boca Raton, FL: CRC Press.
Berger, H. (1929). Uber das electroenkephalogramm de menschem. Archiv fur psychiatric
und Nervenkrankheiten, 97, 527–570.
Binet, A. (1907). The Mind and the Brain. London: K. Paul, Trench, Trübner & Co. Ltd.
Blake, A. J., Rodgers, F. C., Bassuener, A., Hippensteel, J. A., Pearce, T. M., Pearce,
T. R., . . . Williams, J. C. (2010). A microfluidic brain slice perfusion chamber for
multisite recording using penetrating electrodes. Journal of Neuroscience Methods, 189(1),
5–13.
Boahen, K. (2006). Neurogrid: Emulating a million neurons in the cortex. Conference
Proceeding of the IEEE Engineering in Medicine and Biology Society, Supplement, 6702.
Bowers, J. S. (2009). On the biological plausibility of grandmother cells: Implications for
neural network theories in psychology and neuroscience. Psychological Review, 116(1),
220–251.
Breasted, J. H. (1991). The Edwin Smith Surgical Papyrus (Vol. 3–4). Chicago: University
of Chicago Oriental Institute Publications.
Bibliography 185
Brindley, G. S. & Merton, P. A. (1960). The absence of position sense in the human eye.
The Journal of Physiology, 153, 127–130.
Brown, E. N., Kass, R. E., & Mitra, P. P. (2004). Multiple neural spike train data analysis:
State-of-the-art and future challenges. Nature Neuroscience, 7(5), 456–461.
Burks, A. R., & Burks, A. W. (1989). The First Electronic Computer: The Atanasoff Story.
Ann Arbor: University of Michigan Press.
Button, K. S., Ioannidis, J. P. A., Mokrysz, C., Nosek, B. A., Flint, J., Robinson, E. S. J. &
Munafò, M. (2013). Power failure: Why small sample size undermines the reliability
of neuroscience. Nature Reviews Neuroscience, 14, 365–376.
Cajal, S. R. Y. (1906, December 12). The structure and functions of neurons. Paper presented
at the Nobel Foundation Award Ceremony. Stockholm, Sweden.
Carew, T. J., Hawkins, R. D., & Kandel, E. R. (1983). Differential classical-conditioning
of a defensive withdrawal reflex in Aplysia-Californica. Science, 219(4583), 397–400.
Carew, T. J., Pinsker, H. M., & Kandel, E. R. (1972). Long-term habituation of a defensive
withdrawal reflex in Aplysia. Science, 175(4020), 451–454.
Cariani, P. A. (2004). Temporal codes and computations for sensory representation and
scene analysis. IEEE Transactions on Neural Networks, 15(5), 1100–1111.
Carp, J. (2012a). The secret lives of experiments: Methods reporting in the fMRI literature.
NeuroImage, 63(1), 289–300.
Carp, J. (2012b). On the plurality of (methodological) worlds: Estimating the analytic
flexibility of fMRI experiments. Frontiers in Neuroscience, 6, 149.
Carpenter, G. A., Grossberg, S., & Reynolds, J. H. (1991). Artmap—Supervised real-time
learning and classification of nonstationary data by a self-organizing neural network.
Neural Networks, 4(5), 565–588.
Castellucci, V., & Kandel, E. R. (1976). Presynaptic facilitation as a mechanism for behav-
ioral sensitization in Aplysia. Science, 194(4270), 1176–1178.
Castellucci, V., Pinsker, H., Kupferma, I., & Kandel, E. R. (1970). Neuronal mechanisms
of habituation and dishabituation of gill-withdrawal reflex in Aplysia. Science, 167(3926),
1745–1748.
Chalmers, D. J. (1996). The Conscious Mind: In Search of a Fundamental Theory. New York:
Oxford University Press.
Charkhkar, H., Knaack, G. L., Gnade, B. E., Keefer, E. W., & Pancrazio, J. J. (2012).
Development and demonstration of a disposable low-cost microelectrode array for
cultured neuronal network recording. Sensors and Actuators B-Chemical, 161(1),
655–660.
Chiang, A. S., Lin, C. Y., Chuang, C. C., Chang, H. M., Hsieh, C. H., Yeh, C. W., . . .
Hwang, J. K. (2011). Three-dimensional reconstruction of brain-wide wiring networks
in Drosophila at single-cell resolution. Current Biology, 21(1), 1–11.
Christofides, N. (1975). Graph Theory: An Algorithmic Approach. New York: Academic
Press.
Cogan, S. F. (2008). Neural stimulation and recording electrodes. Annual Review of Bio-
medical Engineering, 10, 275–309.
Cohen, T. E., Kaplan, S. W., Kandel, E. R., & Hawkins, R. D. (1997). A simplified prepa-
ration for relating cellular events to behavior: Mechanisms contributing to habituation,
dishabituation, and sensitization of the Aplysia gill-withdrawal reflex. Journal of Neu-
roscience, 17(8), 2886–2899.
Coltheart, M. (2006). What has functional neuroimaging told us about the mind (so far)?
Cortex, 42(3), 323–331.
186 Bibliography
Cox, D. R., & Smith, W. L. (1953). The superposition of several strictly periodic sequences
of events. Biometrika, 40, 1–11.
Cox, D. R., & Smith, W. L. (1954). On the superposition of renewal processes. Biometrika,
41(1–2), 91–99.
Crick, F., & Koch, C. (1995). Are we aware of neural activity in primary visual-cortex?
Nature, 375(6527), 121–123.
Crick, F., & Koch, C. (1998). Consciousness and neuroscience. Cerebral Cortex, 8(2),
97–107.
Cullington, H. E., & Zeng, F. G. (2008). Speech recognition with varying numbers and
types of competing talkers by normal-hearing, cochlear-implant, and implant simula-
tion subjects. Journal of the Acoustical Society of America, 123(1), 450–461.
Damadian, R. (1971). Tumor detection by nuclear magnetic resonance. Science, 171(3976),
1151–1153.
Dawson, G. D., & Scott, J. W. (1949). The recording of nerve action potentials through
skin in man. Journal of Neurology, Neurosurgery, and Psychiatry, 12, 259–267.
de Garis, H., Chen, S., Goertzel, B., & Lian, R. T. (2010). A world survey of artificial
brain projects, Part I: Large-scale brain simulations. Neurocomputing, 74(1–3), 3–29.
DeMonasterio, F. M., & Gouras, P. (1975). Functional properties of ganglion-cells of
rhesus-monkey retina. Journal of Physiology-London, 251(1), 167–195.
Desai, S., Stickney, G., & Zeng, F. G. (2008). Auditory-visual speech perception in normal-
hearing and cochlear-implant listeners. Journal of the Acoustical Society of America, 123(1),
428–440.
DeWeerth, S. P., Meacham, K. W., Giuly, R. J., Guo, L., & Hochman, S. (2008). A
lithographically-patterned, elastic multi-electrode array for surface stimulation of the
spinal cord. Biomedical Microdevices, 10(2), 259–269.
di Pellegrino, G., Fadiga, L., Fogassi, L., Gallese, V., & Rizzolatti, G. (1992). Understanding
motor events—A neurophysiological study. Experimental Brain Research, 91(1),
176–180.
Djourno, A., & Eyries, C. (1957). Prothese auditive par excitation electrique a distance
du nerf sensoriel a laide dun bobinage inclus a demeure. Presse Medicale, 65(63),
1417–1417.
Duhem, P. (1914). La Theorie Physique: Son Objet et Sa Structure. Paris: Marcel Riviera &
Cie.
Ekanadham, C., Tranchina, D., & Simoncelli, E. P. (2014). A unified framework and
method for automatic neural spike identification. Journal of Neuroscience Methods, 222,
47–55.
Eliasmith, C., Stewart, T. C., Choo, X., Bekolay, T., DeWolf, T., Tang, Y., Rasmussen, D.
(2012). A large-scale model of the functioning brain. Science, 338(6111):1202–5.
doi:10.1126/science.1225266
Enroth-Cugell, C., & Robson, J. G. (1966). The contrast sensitivity of retinal ganglion
cells of the cat. Journal of Physiology-London, 187(3), 517–552.
Enserink, M. (2015). A [Euro] billion brain reboot. Science, 347(6229), 1406–1407.
Erickson, J., Tooker, A., Tai, Y. C., & Pine, J. (2008). Caged neuron MEA: A system for
long-term investigation of cultured neural network connectivity. Journal of Neuroscience
Methods, 175(1), 1–16.
Eversmann, B., Jenkner, M., Hofmann, F., Paulus, C., Brederlow, R., Holzapfl, B., . . .
Thewes, R. (2003). A 128x128 CMOS biosensor array for extracellular recording of
neural activity. IEEE Journal of Solid-State Circuits, 38(12), 2306–2317.
Bibliography 187
Farah, M. J. (2014). Brain images, babies, and bathwater: Critiquing critiques of functional
neuroimaging. Hastings Center Report, 44, S19–S30.
Fatt, P., & Katz, B. (1952). Spontaneous subthreshold activity at motor nerve endings.
Journal of Physiology-London, 117(1), 109–128.
Feigl, H. (1958). The mental and the physical. Minnesota Studies in the Philosophy of
Science, 2, 370–497.
Field, D. J. (1994). What is the goal of sensory coding? Neural Computation, 6(4), 559–601.
Flood, T. F., Gorczyca, M., White, B. H., Ito, K., & Yoshihara, M. (2013). A large-scale
behavioral screen to identify neurons controlling motor programs in the Drosophila
brain. G3-Genes Genomes Genetics, 3(10), 1629–1637.
Flynn, M. J. (1972). Some computer organizations and their effectiveness. IEEE Transac-
tions on Computers, C-21(9), 948–960.
Fodor, J. A. & Pylyshyn, Z. W. (1988). Connectionism and cognitive architecture: A
critical analysis. Cognition, 28, 3–71.
Fox, M. D., Snyder, A. Z., Vincent, J. L., Corbetta, M., Van Essen, D. C., & Raichle,
M. E. (2005). The human brain is intrinsically organized into dynamic, anticorrelated
functional networks. Proceedings of the National Academy of Sciences of the United States
of America, 102(27), 9673–9678.
Fox, P. T., & Friston, K. J. (2012). Distributed processing: Distributed functions? Neuro-
Image, 61(2), 407–426.
Fried, I., MacDonald, K. A., & Wilson, C. L. (1997). Single neuron activity in human
hippocampus and amygdala during recognition of faces and objects. Neuron, 18(5),
753–765.
Galambos, R. (1996). Autobiography. In L. Squires (Ed.), The History of Neuroscience in
Autobiography (Vol. 1, pp. 178–221). Washington, DC: Society for Neuroscience.
Galambos, R. (2007). Reflections on the conceptual origins of neuron-glia interactions.
Neuron Glia Biology, 3, 89–91.
Gall, F. J., & Spurzheim, J. C. (1808). Recherches sur le systeme nerveux en general et sur
celui du cerveau en particulier. Paris: Memoirs de Academie de Sciences.
Gallese, V., Fadiga, L., Fogassi, L., & Rizzolatti, G. (1996). Action recognition in the pre-
motor cortex. Brain, 119, 593–609.
Gerard, R. W., & Forbes, A. (1928). “Fatigue” of the flexion reflex. American Journal of
Physiology, 86(1), 186–205.
Gerstein, G. L., & Clark, W. A. (1964). Simultaneous studies of firing patterns in several
neurons. Science, 143(361), 1325–1357.
Gill, S. (1958). Parallel programming. Computer Journal, 1(1), 2–10.
Gingrich, K. J., Baxter, D. A., & Byrne, J. H. (1988). Mathematical-model of cellular
mechanisms contributing to presynaptic facilitation. Brain Research Bulletin, 21(3),
513–520.
Gingrich, K. J., & Byrne, J. H. (1985). Simulation of synaptic depression, posttetanic poten-
tiation, and presynaptic facilitation of synaptic potentials from sensory neurons mediating
gill-withdrawal reflex in Aplysia. Journal of Neurophysiology, 53(3), 652–669.
Goertzel, B., Lian, R. T., Arel, I., de Garis, H., & Chen, S. (2010). A world survey of
artificial brain projects, Part II: Biologically inspired cognitive architectures. Neurocom-
puting, 74(1–3), 30–49.
Goldstein, H. H., & Goldstein, A. (1946). The ENIAC. M.T.A.C., 2, 97–110.
Golgi, C. (1906). The neuron doctrine—theory and facts. Paper presented at the 1906 Nobel
Foundation Awards Ceremony. Stockholm, Sweden.
188 Bibliography
Gonzalez Castillo, J., Saad, Z. S., Handwerker, D. A., Inati, S. J., Brenowitz, N., & Ban-
dettini, P. A. (2012). Whole-brain, time-locked activation with simple tasks revealed
with massive averaging and model-free analysis. Proceedings of the National Academy of
Sciences of the United States of America, 109, 5487–5492.
Goodman, M. B., Hall, D. H., Avery, L., & Lockery, S. R. (1998). Active currents regulate
sensitivity and dynamic range in C. elegans neurons. Neuron, 20(4), 763–772.
Graham, J., & Gerard, R. W. (1946). Journal of Cellular and Comparative Physiology, 28, 99.
Gross, C. G. (2002). Genealogy of the “grandmother cell.” Neuroscientist, 8(5),
512–518.
Gross, G. W., Rieske, E., Kreutzberg, G. W., & Meyer, A. (1977). New fixed-array multi-
microelectrode system designed for long-term monitoring of extracellular single unit
neuronal-activity in vitro. Neuroscience Letters, 6(2–3), 101–105.
Gross, G. W., Williams, A. N., & Lucas, J. H. (1982). Recording of spontaneous activity
with photoetched microelectrode surfaces from mouse spinal neurons in culture.
Journal of Neuroscience Methods, 5(1–2), 13–22.
Grossberg, S., & Mingolla, E. (1985). Neural dynamics of perceptual grouping: Textures,
boundaries, and emergent segmentations. Perception & Psychophysics, 38(2), 141–171.
Grundfest, H., Sengstaken, R. W., Oettinger, W. H., & Gurry, R. W. (1950). Stainless steel
micro-needle electrodes made by electrolytic pointing. Review of Scientific Instruments,
21(4), 360–361.
Ha, H. I., Hendricks, M., Shen, Y., Gabel, C. V., Fang-Yen, C., Qin, Y. Q., . . . Zhang, Y.
(2010). Functional organization of a neural network for aversive olfactory learning in
Caenorhabditis elegans. Neuron, 68(6), 1173–1186.
Ham, F. M., & Kostanic, I. (2000). Principles of Neurocomputing in Science and Engineering.
New York: McGraw-Hill.
Harary, F. (1969). Graph Theory. Reading, MA: Addison-Wesley Pub. Co.
Harris, J. D. (1943). Habituatory response decrement in the intact organism. Psychological
Bulletin, 40(6), 385–422.
Hartline, H. K. (1938). The response of single optic nerve fibers of the vertebrate eye to
illumination of the retina. American Journal of Physiology, 121(2), 400–415.
Hartline, H. K. (1940). The effects of spatial summation in the retina on the excitation
of the fibers of the optic nerve. American Journal of Physiology, 130(4), 700–711.
Hartline, H. K. (1949). Inhibition of activity of visual receptors by illuminating nearby
retinal areas in the Limulus eye. Federation Proceedings, 8(1), 69–69.
Hartline, H. K., & Graham, C. H. (1932). Nerve impulses from single receptors in the
eye. Journal of Cellular and Comparative Physiology, 1(2), 277–295.
Hartline, H. K., & Ratliff, F. (1957). Inhibitory interaction of receptor units in the eye
of Limulus. Journal of General Physiology, 40, 357–376.
Hartline, H. K., & Ratliff, F. (1958). Spatial summation of inhibitory influences in the
eye of Limulus, and the mutual interaction of receptor units. Journal of General Physiol-
ogy, 41, 1049–1066.
Hartline, H. K., Wagner, H. G., & Ratliff, F. (1956). Inhibition in the eye of Limulus.
Journal of General Physiology, 39(5), 651–673.
Hebb, D. O. (1949). The Organization of Behavior: A Neuropsychological Theory. New York:
Wiley.
Heer, F., Franks, W., Blau, A., Taschini, S., Ziegler, C., Hierlemann, A., & Baltes, H. (2004).
CMOS microelectrode array for the monitoring of electrogenic cells. Biosensors &
Bioelectronics, 20(2), 358–366.
Bibliography 189
Hell, S. W., & Wichmann, J. (1994). Breaking the diffraction resolution limit by stimulated-
emission—Stimulated-emission-depletion fluorescence microscopy. Optics Letters,
19(11), 780–782.
Henson, R. (2006). Forward inference using functional neuroimaging: Dissociations versus
associations. TRENDS in Cognitive Sciences, 10(2), 64–69.
Hertz, J., Krogh, A., & Palmer, R. (1991). Introduction to the Theory of Neural Computation.
Redwood City, CA: Addison-Wesley.
Hickok, G. (2009). Eight problems for the mirror neuron theory of action understanding
in monkeys and humans. Journal of Cognitive Neuroscience, 21(7), 1229–1243.
Hilgetag, C. C., O’Neill, M. A., & Young, M. P. (1996). Indeterminate organization of
the visual system—Rejoinder. Science, 271(5250), 777–777.
Hinton, G. E., & Sejnowski, T. J. (1986). Learning and relearning in Boltzmann machines.
In D. E. Rumelhart & J. L. McClelland (Eds.), Parallel Distributed Processing. Explora-
tions in the Microstructure of Cognition. Vol. 1: Foundations. Cambridge, MA: MIT Press.
Hodgkin, A. L., & Huxley, A. F. (1952). Currents carried by sodium and potassium ions
through the membrane of the giant axon of Loligo. Journal of Physiology, 116(4),
449–472.
Hodgkin, A. L., Huxley, A. F., & Katz, B. (1952). Measurement of current-voltage rela-
tions in the membrane of the giant axon of Loligo. Journal of Physiology, 116(4),
424–448.
Hodgkin, A. L., & Katz, B. (1949). The effect of sodium ions on the electrical activity
of giant axon of the squid. Journal of Physiology, 108(1), 37–77.
Hoogerwerf, A. C., & Wise, K. D. (1994). A 3-dimensional microelectrode array for
chronic neural recording. IEEE Transactions on Biomedical Engineering, 41(12),
1136–1146.
Hopfield, J. J. (1982). Neural networks and physical systems with emergent collective
computational abilities. Proceedings of the National Academy of Sciences of the United States
of America-Biological Sciences, 79(8), 2554–2558.
Horvath, J. C., Carter, O., & Forte, J. D. (2014). Transcranial direct current stimulation:
Five important issues we aren’t discussing (but probably should be). [Review]. Frontiers
in Systems Neuroscience, 8, 2.
Horvath, J. C., Forte, J. D., & Carter, O. (2015a). Quantitative review finds no evidence
of cognitive effects in healthy populations from single-session transcranial; direct cur-
rent stimulation (tDCS). Brain Stimulation, 8(3), 535–550.
Horvath, J. C., Forte, J. D., & Carter, O. (2015b). Evidence that transcranial direct cur-
rent stimulation (tDCS) generates little-to-no reliable neurophysiologic effect beyond
MEP amplitude modulation in healthy human subjects: A systematic review. Neuro-
psychologia, 66, 213–236.
Horwitz, B. (2003). The elusive concept of brain connectivity. NeuroImage, 19(2),
466–470.
Hubel, D. H. (1957). Tungsten microelectrode for recording from single units. Science,
125(3247), 549–550.
Hubel, D. H., & Wiesel, T. N. (1959). Receptive fields of single neurones in the cat’s
striate cortex. The Journal of Physiology, 148, 574–571.
Hubel, D. H., & Wiesel, T. N. (1965). Receptive fields and functional architecture in 2
nonstriate visual areas (18 and 19) of cat. Journal of Neurophysiology, 28(2), 229–289.
Hubel, D. H., & Wiesel, T. N. (1968). Receptive fields and functional architecture of
monkey striate cortex. Journal of Physiology-London, 195(1), 215–243.
190 Bibliography
Maccione, A., Garofalo, M., Nieus, T., Tedesco, M., Berdondini, L., & Martinoia, S. (2012).
Multiscale functional connectivity estimation on low-density neuronal cultures recorded
by high-density CMOS Micro Electrode Arrays. Journal of Neuroscience Methods, 207(2),
161–171.
MacCorquodale, K., & Meehl, P. E. (1948). On a distinction between hypothetical con-
structs and intervening variables. Psychological Review, 55(2), 95–107.
MacNichol, E. F., Jr. (1956). Visual receptors as biological transducers. In R. G. Grenell &
L. J. Mullins (Eds.), Molecular Structure and Functional Activity of Nerve Cells (Publication
1, 34). Washington, D.C.: American Institute of Biological Sciences.
Margolis, D. J., Gartland, A. J., Singer, J. H., & Detwiler, P. B. (2014). Network oscil-
lations drive correlated spiking of ON and OFF ganglion cells in the rd1 mouse model
of retinal degeneration. PLoS One, 9(1), e86253. doi:10.1371/journal.pone.0086253
Markram, H. (2006). The Blue Brain Project. Nature Reviews Neuroscience, 7(2), 153–160.
Martin, S. J., Grimwood, P. D., & Morris, R. G. (2000). Synaptic plasticity and memory:
An evolution of the hypothesis. Annual Review Neuroscience, 23, 649–711.
McClelland, J. L., & Rumelhart, D. E. (1988). A simulation-based tutorial system for
exploring parallel distributed-processing. Behavior Research Methods Instruments & Com-
puters, 20(2), 263–275.
McClelland, J. L., Rumelhart, D. E., & The PDP Research Group (1988). Parallel Dis-
tributed Processing: Explorations in the Microstructure of Cognition. Vol 2 Psychological and
Biological Models. Cambridge, MA: MIT Press.
McCulloch, W. S. (1944). The functional organization of the cerebral cortex. Physiological
Reviews, 24(3), 390–407.
McCulloch, W. S., & Pitts, W. (1943). A logical calculus of the ideas immanent in nervous
activity. Bulletin of Mathematical Biophysics, 5, 115–133.
Merolla, P. A., Arthur, J. V., Alvarez-Icaza, R., Cassidy, A. S., Sawada, J., Akopyan, F., . . .
Modha, D. S. (2014). Artificial brains: A million spiking-neuron integrated circuit
with a scalable communication network and interface. Science, 345(6197), 668–673.
Meyer, A. R. (1974). The inherent computational complexity of theories of ordered sets. Paper
presented at the Proceedings of the International Congress of Mathematicians, Van-
couver, Canada.
Minsky, M. L., & Papert, S. (1969). Perceptrons: An Introduction to Computational Geometry.
Cambridge, MA: MIT Press.
Moore, E. F. (1956). Gedanken—experiments on sequential machines. In C. E. Shannon &
J. McCarthy (Eds.), Automata Studies (pp. 129–153). Princeton, NJ: Princeton Uni-
versity Press.
Moore, G. E. (1965). Cramming more components onto integrated circuits. Electronics,
38(8), 114–117.
Moxon, K. A., Hallman, S., Sundarakrishnon, A., Wheatley, M., Nissanov, J., & Barbee,
K. A. (2009). Long-term recordings of multiple single-neurons for clinical applications:
The emerging role of the bioactive microelectrode. Materials, 2, 1762–1794.
Mukamel, D., Ruffo, S., & Schreiber, N. (2005). Breaking of ergodicity and long relaxation
times in systems with long-range interactions. Physical Review Letters, 95(24), 240604.
Murphy, T. H., Blatter, L. A., Wier, W. G., & Baraban, J. M. (1992). Spontaneous syn-
chronous synaptic calcium transients in cultured cortical-neurons. Journal of Neuroscience,
12(12), 4834–4845.
Nabavi, S., Fox, R., Perolux, C. D., Lin, J. Y., Tsien, R. Y., & Malinow, R. (2014). Engi-
neering a memory with LTD and LTP. Nature, 511, 348–352.
Bibliography 193
Neri, P., & Levi, D. M. (2006). Receptive versus perceptive fields from the reverse-correlation
viewpoint. Vision Res, 46(16), 2465–2474.
Nicolelis, M. A. L., & Chapin, J. K. (1994). Spatiotemporal structure of somatosensory
responses of many-neuron ensembles in the rat ventral posterior medial nucleus of
the thalamus. Journal of Neuroscience, 14(6), 3511–3532.
Nordhausen, C. T., Maynard, E. M., & Normann, R. A. (1996). Single unit recording
capabilities of a 100 microelectrode array. Brain Research, 726(1–2), 129–140.
Ogawa, S., Lee, T. M., Kay, A. R., & Tank, D. W. (1990). Brain magnetic-resonance-
imaging with contrast dependent on blood oxygenation. Proceedings of the National
Academy of Sciences of the United States of America, 87(24), 9868–9872.
Olsen, S. R., & Wilson, R. I. (2008). Cracking neural circuits in a tiny brain: New
approaches for understanding the neural circuitry of Drosophila. Trends in Neurosciences,
31(10), 512–520.
Oosterwijk, S., Lindquist, K. A., Anderson, E., Dautoff, R., Moriguchi, Y., & Bennett, L.
F. (2012). States of mind: Emotions, body feelings, and thoughts share distributed
neural networks. NeuroImage 62, 2110–2128.
Palmer, S. E. (1999). Vision science: Photons to phenomenology. Cambridge, MA: MIT.
Papadimitriou, C. H., & Tsitsiklis, J. (1986). Intractable problems in control-theory. Siam
Journal on Control and Optimization, 24(4), 639–654.
Perrett, D. I., Rolls, E. T., & Caan, W. (1982). Visual neurones responsive to faces in the
monkey temporal cortex. Experimental Brain Research, 47(3), 329–342.
Pickel, V. M., & Segal, M. (Eds.). (2013). The Synapse: Structure and Function. Oxford, UK:
Academic Press.
Pillow, J. W., Shlens, J., Chichilnisky, E. J., & Simoncelli, E. P. (2013). A model-based
spike sorting algorithm for removing correlation artifacts in multi-neuron recordings.
PLoS One, 8(5), e62123. doi:10.1371/journal.pone.0062123
Pine, J. (1980). Recording action-potentials from cultured neurons with extracellular
micro-circuit electrodes. Journal of Neuroscience Methods, 2(1), 19–31.
Pinsker, H., Kupfermann, I., Castellucci, V., & Kandel, E. (1970). Habituation and disha-
bituation of the gill withdrawal reflex in Aplysia. Science, 167, 1740–1732.
Pitts, W., & McCulloch, W. S. (1947). How we know universals: The perception of audi-
tory and visual forms. Bulletin of Mathematical Biophysics, 9, 127–147.
Place, U. T. (1956). Is consciousness a brain process? British Journal of Psychology, 47(1),
44–50.
Poldrack, R. A. (2006). Can cognitive processes be inferred from neuroimaging data?
Trends in Cognitive Sciences, 10(2), 59–63.
Poldrack, R. A. (2010). Mapping mental function to brain structure: How can cognitive
neuroimaging succeed? Perspectives on Psychological Science, 5(6), 753–761.
Posner, M. I., Petersen, S. E., Fox, P. T., & Raichle, M. E. (1998). Localization of cognitive
operations in the human brain. Science, 240, 1627–1631.
Quine, W. V. O. (1951). Main trends in recent philosophy. Philosophical Review, 60, 20–43.
Quiroga, R. Q. (2012). Concept cells: The building blocks of declarative memory func-
tions. Nature Reviews Neuroscience, 13(8), 587–597.
Quiroga, R. Q., Nadasdy, Z., & Ben-Shaul, Y. (2004). Unsupervised spike detection and
sorting with wavelets and superparamagnetic clustering. Neural Computation, 16(8),
1661–1687.
Quiroga, R. Q., Reddy, L., Kreiman, G., Koch, C., & Fried, I. (2005). Invariant visual
representation by single neurons in the human brain. Nature, 435(7045), 1102–1107.
194 Bibliography
Rabi, I. I., Zacharias, J. R., Millman, S., & Kusch, P. (1937). A new method for measuring
nuclear magnetic moments. Physical Review, 53, 318.
Rabi, I. I., Zacharias, J. R., Millman, S., & Kusch, P. (1939). The molecular beam resonance
method for measuring nuclear magnetic moments. The magnetic moments of 3LI6,
3LI7, and 9F19. Physical Review, 55, 526–535.
Raizen, D. M., & Avery, L. (1994). Electrical-activity and behavior in the pharynx of C.
elegans. Neuron, 12(3), 483–495.
Rakover, S. S. (2011). A plea for methodological dualism and a multi-explanation approach
in psychology. Behavior and Philosophy, 39, 17–43.
Ratliff, F., Hartline, H. K., & Miller, W. H. (1963). Spatial and temporal aspects of retinal
inhibitory interaction. Journal of the Optical Society of America, 53, 110–120.
Richmond, J. E. (2006). Electrophysiological recordings from the neuromuscular junction
of C. elegans. In The C. elegans Research Community (Ed.), WormBook, 1–8.
Rosenblatt, F. (1958). The perceptron—A probabilistic model for information-storage and
organization in the brain. Psychological Review, 65(6), 386–408.
Roskies, A. L. (2010). How does neuroscience affect our conception of volition? Annual
Review of Neuroscience, 33, 190–130.
Rousche, P. J., & Normann, R. A. (1998). Chronic recording capability of the Utah
intracortical electrode array in cat sensory cortex. Journal of Neuroscience Methods, 82(1),
1–15.
Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1986). Learning representations by
back-propagating errors. Nature, 323(6088), 533–536.
Rumelhart, D. E., McClelland, J. L., & The PDP Research Group (1988). Parallel Dis-
tributed Processing. Explorations in the Microstructure of Cognition. Vol 1: Foundations.
Cambridge, MA: MIT Press.
Schwarz, D. A., Lebedev, M. A., Hanson, T. L., Dimitrov, D. F., Lehew, G., Meloy, J., . . .
Nicolelis, M. A. (2014). Chronic, wireless recordings of large-scale brain activity in
freely moving rhesus monkeys. Nature Methods, 11, 670–676.
Searle, J. R. (1990). The Mystery of Consciousness. New York: New York Review of
Books.
Searle, J. R. (1997, March 6). Consciousness and the Philosophers. New York: New York
Review of Books.
Sejnowski, T. J., Kienker, P. K., & Hinton, G. E. (1986). Learning symmetry groups with
hidden units—Beyond the perceptron. Physica D-Nonlinear Phenomena, 22(1–3),
260–275.
Seker, E., Berdichevsky, Y., Begley, M. R., Reed, M. L., Staley, K. J., & Yarmush, M. L.
(2010). The fabrication of low-impedance nanoporous gold multiple-electrode arrays
for neural electrophysiology studies. Nanotechnology, 21(12). doi:10.1088/0957-4484/
21/12/125504
Selfridge, O. G. (1958). Pandemonium: A paradigm for learning. Paper presented at the
Proceeding of a Symposium the Mechanization of Thought Processes, London.
Service, R. F. (2014). The brain chip. Science, 345(6197), 614–616.
Sevush, S. (2006). Single-neuron theory of consciousness. Journal of Theoretical Biology,
238(3), 704–725.
Sherrington, C. S. (1940). Man on His Nature. Cambridge, UK: The University Press.
Shlens, J., Field, G. D., Gauthier, J. L., Greschner, M., Sher, A., Litke, A. M., & Chich-
ilnisky, E. J. (2009). The structure of large-scale synchronized firing in primate retina.
Journal of Neuroscience, 29(15), 5022–5031.
Bibliography 195
Simmons, J. P., Nelson, L. D., & Simonsohn, U. (2011). False-positive psychology: Undis-
closed flexibility in data collection and analysis allows presenting anything as significant.
Psychological Science, 22(11), 1359–1366.
Skinner, B. F. (1950). Are theories of learning necessary? Psychological Review, 57,
193–216.
Spillmann, L. (2014). Receptive fields of visual neurons: The early years. [Research Sup-
port, Non-U.S. Gov’t]. Perception, 43(11), 1145–1176.
Spillmann, L., Dresp-Langley, B., & Tseng, C.-H. (2015). Beyond the classical receptive
field: The effect of contextural stimuli. Journal of Vision, 15(9), 7, 1–23.
Spillmann, L., & Werner, J. S. (Eds.). (1990). Visual Perception: The Neurophysiological
Foundations. San Diego: Academic Press.
Spira, M. E., & Hai, A. (2013). Multi-electrode array technologies for neuroscience and
cardiology. Nature Nanotechnology, 8(2), 83–94.
Sporns, O. (2011). Networks of the Brain. Cambridge, MA: MIT Press.
Sporns, O., Tononi, G., & Kotter, R. (2005). The human connectome: A structural descrip-
tion of the human brain. PLoS Computational Biology, 1(4), 245–251.
Steidl, E. M., Neveu, E., Bertrand, D., & Buisson, B. (2006). The adult rat hippocampal
slice revisited with multi-electrode arrays. Brain Research, 1096, 70–84.
Stevenson, I. H., Rebesco, J. M., Hatsopoulos, N. G., Haga, Z., Miller, L. E., & Kording,
K. P. (2009). Bayesian inference of functional connectivity and network structure from
spikes. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 17(3),
203–213.
Stockmeyer L. J., & Chandra, A. K. (1979). Provably difficult combinatorial games, SIAM
Journal on Computing, 8(2), 151–174.
Takeuchi, T., Duszkiewicz, A. J., & Morris, R. G. (2013). The synaptic plasticity and
memory hypothesis: Encoding, storage and persistence. [Research Support, Non-U.S.
Gov’t]. Philosophical Transactions of the Royal Society of London Series B-Biological Sciences,
369(1633), 20130288.
Tanaka, K. (1993). Neuronal mechanisms of object recognition. Science, 262(5134),
685–688.
Thomas, C. A., Jr., Springer, P. A., Loeb, G. E., Berwald-Netter, Y., & Okun, L. M.
(1972). A miniature microelectrode array to monitor the bioelectric activity of cultured
cells. Experimental Cell Research, 74(1), 61–66.
Thompson, R. F., & Spencer, W. A. (1966). Habituation—A model phenomenon for
study of neuronal substrates of behavior. Psychological Review, 73(1), 16–43.
Thyreau, B., Schwartz, Y., Thirionm, B., Frouin, V., Loth, E., Vollstädt-Klein, S., . . .
IMAGEN Consortium. (2012). Very large fMRI study using the IMAGEN database:
Sensitivity-specificity and population effect modeling in relation to the underlying
anatomy. NeuroImage, 61(1), 295–303.
Turella, L., Pierno, A. C., Tubaldi, F., & Castiello, U. (2009). Mirror neurons in humans:
Consisting or confounding evidence? Brain and Language, 108(1), 10–21.
Turing, A. (1950). Computing machinery and intelligence. Mind, 59, 433–460.
Turing, A. (2004). Lecture on the automatic computing engine. In B. J. Copeland (Ed.),
The Essential Turing (pp. 362–394). Oxford: Oxford University Press.
Turkeltaub, P. E., Eden, G. F., Jones, K. M., & Zeffiro, T. A. (2002). Meta-analysis of the
functional neuroanatomy of single-word reading: Method and validation. NeuroImage,
16(3), 765–780.
Underwood, E. (2014). Brain Project meets physics. Science, 344, 954–955.
196 Bibliography
Ward, A. A., & Thomas, L. B. (1955). The electrical activity of single units in the cerebral
cortex of man. Electroencephalography and Clinical Neurophysiology, 7(1), 135–136.
Wark, H. A., Sharma, R., Mathews, K. S., Fernandez, E., Yoo, J., Christensen, B., Tresco,
P. . . . Tathireddy, P. (2013). A new high-density (25 electrodes/mm²) penetrating
microelectrode array for recording and stimulating sub-millimeter neuroanatomical
structures. Journal of Neural Engineering, 10(4):045003. doi:10.1088/1741-2560/
10/4/045003.
Watson, J. D., & Crick, F. H. C. (1953). Molecular structure of nucleic acids—A structure
for deoxyribose nucleic acid. Nature, 171(4356), 737–738.
Westheimer, G. (1990). Relating neural mechanisms to visual perception. In L. Spillman &
J. S. Werner (Eds.), Visual Perception: The Neurophysiological Foundations. San Diego:
Academic Press.
White, J. G., Southgate, E., Thomson, J. N., & Brenner, S. (1986). The structure of the
nervous-system of the nematode C. elegans. Philosophical Transactions of the Royal Society
of London Series B-Biological Sciences, 314(1165), 1–340.
Wild, J., Prekopcsak, Z., Sieger, T., Novak, D., & Jech, R. (2012). Performance comparison
of extracellular spike sorting algorithms for single-channel recordings. Journal of Neu-
roscience Methods, 203(2), 369–376.
Wilson-Sanders, S. E. (2011). Invertebrate models for biomedical research, testing, and
education. ILAR Journal, 52(2), 126–152.
Yoshimura, Y., Dantzker, J. L. M., & Callaway, E. M. (2005). Excitatory cortical neurons
form fine-scale functional networks. Nature, 433(7028), 868–873.
Young, J. Z. (1936). The giant nerve fibres and epistellar body of cephalopods. Quarterly
Journal of Microscopical Science, 78, 367–386.
Young, M. P., & Yamane, S. (1992). Sparse population coding of faces in the inferotem-
poral cortex. Science, 256(5061), 1327–1331.
Yule, G. U. (1926). Why do we sometimes get nonsense-correlations between time-series?—
A study in sampling and the nature of time-series. Journal of the Royal Statistical Society,
89, 1–69.
Zecevic, D., Wu, J. Y., Cohen, L. B., London, J. A., Hopp, H. P., & Falk, C. X. (1989).
Hundreds of neurons in the Aplysia abdominal-ganglion are active during the gill-
withdrawal reflex. Journal of Neuroscience, 9(10), 3681–3689.
Zeng, F. G., Rebscher, S., Harrison, W., Sun, X., & Feng, H. (2008). Cochlear implants:
System design, integration, and evaluation. [Research Support, N.I.H., Extramural
Research Support, Non-U.S. Gov’t Review]. IEEE Reviews in Biomedical Engineering,
1, 115–142.
Zhang, Y. L., Liu, R. Y., Heberton, G. A., Smolen, P., Baxter, D. A., Cleary, L. J., & Byrne,
J. H. (2012). Computational design of enhanced learning protocols. Nature Neurosci-
ence, 15(2), 294–297.
Zrenner, E. (2002). Will retinal implants restore vision? [Research Support, Non-U.S.
Gov’t Review]. Science, 295(5557), 1022–1025.
This page intentionally left blank
INDEX
Note: Figures and tables are denoted with italicized page numbers; endnote information is
denoted with an n and note number following the page number.