Choose Which Theory You Prefer
Computational Models for Neuroscience: Human Cortical Information Processing
Robert Hecht-Nielsen, Thomas McKenna
Formal study of neuroscience (broadly defined) has been underway for millennia. For example, writing
2,350 years ago, Aristotle asserted that association—of which he defined three specific varieties—lies at
the center of human cognition. Over the past two centuries, the simultaneous rapid advancements of
technology and (consequently) per capita economic output have fueled an exponentially increasing
effort in neuroscience research. Today, thanks to the accumulated efforts of hundreds of thousands of
scientists, we possess an enormous body of knowledge about the mind and brain. Unfortunately, much
of this knowledge is in the form of isolated factoids. In terms of "big picture" understanding, surprisingly
little progress has been made since Aristotle. In some arenas we have probably suffered negative
progress because certain neuroscience and neurophilosophy precepts have clouded our self-
knowledge; causing us to become largely oblivious to some of the most profound and fundamental
aspects of our nature (such as the highly distinctive propensity of all higher mammals to automatically
segment all aspects of the world into distinct holistic objects and the massive reorganization of large
portions of our brains that ensues when we encounter completely new environments and life situations).
At this epoch, neuroscience is like a huge collection of small, jagged, jigsaw puzzle pieces piled in a
mound in a large warehouse (with neuroscientists going in and tossing more pieces onto the mound
every month). Few attempts to find pieces in the mound that fit together have succeeded. Based upon
this record of failure, few such attempts are now launched. Neuroscience research today is largely
centered on the activity of uncovering new factoids in a careful, systematic fashion and then tossing
them on the mound.
Among neuroscientists there is a general anticipation that someday (not soon) somebody will fit enough
pieces of the puzzle together to discern a key concept that will trigger rapid and sustained progress in
assembling the rest of the existing pieces and in discovering the still missing pieces. The purpose of this
book is to try to hurry up this timetable by describing several different conceptual frameworks that
indicate how some of the pieces of the neuroscience puzzle may fit together. Perhaps the trigger
concept that will lead to the rapid assembly of the neuroscience jigsaw puzzle is in this book. Or, if not,
perhaps this book will inspire creation of the trigger concept.
Chapter 1 of this book, by Larry Cauller (the chapters are organized alphabetically by author name),
presents a dynamics approach to understanding cortical function. As with many of the models and
constructs presented here, Cauller's theory has been computer implemented and has been shown to
have interesting and relevant properties. In particular, he concludes that it is the dynamical attractors of
neuronal networks which are the fundamental, repeatable, basic elements of information storage and
processing. Ideas such as neurons functioning as simple "feature detectors" are rejected. The net result
is an iconoclast and fresh viewpoint about the workings of cerebral cortex.
Chapter 2, by Oleg Favorov and colleagues, takes a radically different viewpoint in comparison with
Chapter 1; namely, that Frank Rosenblatt had it right all along. In this view, the brain is composed of
perceptron-like multilayer neuron networks trained by backpropagation. Favorov's theory is well
argued. The main point is that this concept jibes with many known properties of cerebral information
processing and behavior. The fact that "connectionist" cognitive science has been astoundingly
successful in building phenomenological models of many cortical information processing operations is
another strong argument for this theory.
Chapter 3, by Walter Freeman, is philosophically in tune with Chapter 1, but takes the main point even
further, invoking chaos as a key ingredient in brain dynamics. In Freeman's conception, each neuron
participates in multiple, interlinked, dynamical interactions with other neurons. Going farther than
Cauller, Freeman's theory views trajectories transitioning from one attractor to another as the
representations of information. The learning and behaving processes hinge on the occurrences of these
transitions, not on the sustained attainment of stable attractor limit cycles. Appropriate (goal-satisfying)
transitions are reinforced and enhanced, neutral or goal-denying transitions are suppressed. Freeman's
viewpoint has heft because its main tenets are behaviors that real neurons can probably achieve and
because it helps explain why conventional concepts have had such little success in explaining brain
information processing.
Chapter 4, by Robert Hecht-Nielsen, presents an almost entirely new, detailed and complete, high-level
theory of thalamocortical information processing. It postulates that all thalamocortical information
processing can be accounted for by two particular types of associative memory neural networks, each
employed in abundance. The first of these networks is a fast-acting two-step attractor system
(implemented by reciprocal connections between cortex and thalamus) which casts all information into
one of a few thousand fixed-for-life sparse neuronal codes (representational tokens). The cerebral
cortex is presumed to be tiled with tens of thousands of these (only weakly interacting) networks. The
second network learns to connect certain pairs of these tokens on different regions. The strength of
connection is a novel "fuzzy" information theoretic quantity called antecedent support; which Hecht-
Nielsen claims is the only information learned, stored, and used in the thalamocortex. The notion that
the cortex is fundamentally symbolic may appeal to old-school Al researchers. "Computational
Intelligence" fans may like the theory because the learned relationships between pairs of hard symbols
are soft and yet information-theoretic in character. Neuroscientists may like it because the whole thing
seems to exactly match the known anatomy and physiology of the thalamocortex and its related
structures.
Chapter 5, by Henry Markram, is a tour de force of advanced insight into those most critical brain
elements: synapses. Markram's synthesis of the huge and confusing literature on this subject is guided
by his own world-leading experimental findings. In particular, in strong resonance with several other
chapters of the book, he emphasizes the central importance of dynamics in synaptic transmission: the
post-synaptic effect of an arriving action potential depends critically on the recent history of arrivals of
other action potentials. Going far beyond a mere recitation of the facts, Markram presents a compelling
synthesis that identifies and characterizes principal categories of synaptic dynamical behavior. This
chapter provides abundant knowledge of great value to all interested in cortical theory.
Chapter 6, by Thomas McKenna, provides sage commentary on a variety of cortical theories and their
relationships. In particular, McKenna emphasizes and explores the gaps that clearly exist between
cortical theories and biology. These gaps are then used as pointers towards potentially profitable new
directions of theory construction. A valuable feature of McKenna's chapter is his characterizations and
comparisons of theories that operate at different levels of abstraction. Going far beyond the simple "top
down" and "bottom up" taxonomy, theories are discussed in terms of their ability to succeed at their
individual level of abstraction. Two hundred years from now any high school student will probably be
able to give a simple explanation for how the brain represents objects, learns and stores facts about
those objects and then uses those facts to carry out information processing. A key point is that a correct
theory can function without gaps at any level of detail. Another issue addressed in this chapter is the
applicability of cortical theories. For theories to be judged correct and complete they must support
practical implementations which can solve high-value human problems.
Chapter 7, by Jeffrey Sutton and Gary Strangman, proposes that the cerebral cortex can be viewed as
being composed of hierarchical groups of neural networks; or, as they term it, a Network of Networks
(NoN). The main point of this theory is that only modeling individual associative memory networks will
never, by itself, explain the function of the cerebral cortex. A key missing ingredient is that groups of
such networks are used, controlled, and trained as an ensemble—not individually. The chapter then
goes on to explain how this might be accomplished and why this view is so important in understanding
the obvious hierarchicality of processing we are so intimately familiar with (our cortex can deal with
"throw the ball" as a conceptual unit just as easily as "slightly increase bicep tension").
Chapter 8, by John Taylor, presents a theory of cerebral information processing that concentrates on
how consciousness may emerge from the combined activity of ensembles of stimulus-driven cortical-
subcortical control structures set up to tend to a variety of behavioral goals and needs. Taylor's analysis
frees itself from the details by fixating on firmly established high-level brain information properties that
any successful theory must explain. Mixing together insights from a variety of theoretical perspectives,
Taylor integrates these into a whole that emerges as a vision of how collections of relatively
straightforward "thermostat" brain functions can, because of the strong causal linkages between their
activating stimuli, end up supporting a higher-level process that becomes self-aware. Unlike past toy-
system-level studies of "emergent" properties, Taylor's theory gets to the heart of why and how
conscious awareness phylogenetically emerged. This chapter illustrates the value, and provides a
masterful example, of theory building at a high conceptual level.
Chapter 9, by Neill Taylor and John Taylor, addresses another key cortex theory issue: language. In
particular, they provide a foundational study of how a neuronal architecture can be developed to
implement Noam Chomsky's hypothetical "Language Acquisition Device." Central to their theory are
structures for chunking and unchunking of word strings (i.e., hierarchical networks). Using these, they
develop a set of neuronally plausible language processing components and procedures that exhibit
many of the key properties that a "LAD" must have. The resulting theory is not intended as a final theory
of language acquisition and understanding. However, it provides a nice jumping-off point for further
research.
Chapter 10, by Richard Zemel, is a theory of the population codes that many researchers believe lie at
the heart of cortical information processing. In the past, almost all studies of population codes have
assumed that the collection of neurons making up such a code are "ON" (and that other nearby neurons
are "OFF"). Even with this binary assumption, many interesting properties of such codes have been
established. In the theory presented in this chapter, Zemel goes way beyond this assumption: he allows
population codes to also have variable non-negative activity levels assigned to them. He then shows
that these activity levels can be interpreted in terms of a related probability distribution. Zemel then
argues that the process of activating a population code is itself a probability calculation. Furthermore,
he shows that collections of linked population codes can interact to form a type of belief network; a belief
network that seems highly compatible with cortical neuronal anatomy, physiology, and function. The net
result is a foundation for a new and promising direction of research.
Which theory do you prefer?
Chapter 1
Chapter 2
Chapter 3
Chapter 4
Chapter 5
Chapter 6
Chapter 7
Chapter 8
Chapter 9
Chapter 10
My own theory
Computational Models for Neuroscience: Human Cortical Information Processing
Robert Hecht-Nielsen, Thomas McKenna
Formal study of neuroscience (broadly defined) has been underway for millennia. For example, writing
2,350 years ago, Aristotle asserted that association—of which he defined three specific varieties—lies at
the center of human cognition. Over the past two centuries, the simultaneous rapid advancements of
technology and (consequently) per capita economic output have fueled an exponentially increasing
effort in neuroscience research. Today, thanks to the accumulated efforts of hundreds of thousands of
scientists, we possess an enormous body of knowledge about the mind and brain. Unfortunately, much
of this knowledge is in the form of isolated factoids. In terms of "big picture" understanding, surprisingly
little progress has been made since Aristotle. In some arenas we have probably suffered negative
progress because certain neuroscience and neurophilosophy precepts have clouded our self-
knowledge; causing us to become largely oblivious to some of the most profound and fundamental
aspects of our nature (such as the highly distinctive propensity of all higher mammals to automatically
segment all aspects of the world into distinct holistic objects and the massive reorganization of large
portions of our brains that ensues when we encounter completely new environments and life situations).
At this epoch, neuroscience is like a huge collection of small, jagged, jigsaw puzzle pieces piled in a
mound in a large warehouse (with neuroscientists going in and tossing more pieces onto the mound
every month). Few attempts to find pieces in the mound that fit together have succeeded. Based upon
this record of failure, few such attempts are now launched. Neuroscience research today is largely
centered on the activity of uncovering new factoids in a careful, systematic fashion and then tossing
them on the mound.
Among neuroscientists there is a general anticipation that someday (not soon) somebody will fit enough
pieces of the puzzle together to discern a key concept that will trigger rapid and sustained progress in
assembling the rest of the existing pieces and in discovering the still missing pieces. The purpose of this
book is to try to hurry up this timetable by describing several different conceptual frameworks that
indicate how some of the pieces of the neuroscience puzzle may fit together. Perhaps the trigger
concept that will lead to the rapid assembly of the neuroscience jigsaw puzzle is in this book. Or, if not,
perhaps this book will inspire creation of the trigger concept.
Chapter 1 of this book, by Larry Cauller (the chapters are organized alphabetically by author name),
presents a dynamics approach to understanding cortical function. As with many of the models and
constructs presented here, Cauller's theory has been computer implemented and has been shown to
have interesting and relevant properties. In particular, he concludes that it is the dynamical attractors of
neuronal networks which are the fundamental, repeatable, basic elements of information storage and
processing. Ideas such as neurons functioning as simple "feature detectors" are rejected. The net result
is an iconoclast and fresh viewpoint about the workings of cerebral cortex.
Chapter 2, by Oleg Favorov and colleagues, takes a radically different viewpoint in comparison with
Chapter 1; namely, that Frank Rosenblatt had it right all along. In this view, the brain is composed of
perceptron-like multilayer neuron networks trained by backpropagation. Favorov's theory is well
argued. The main point is that this concept jibes with many known properties of cerebral information
processing and behavior. The fact that "connectionist" cognitive science has been astoundingly
successful in building phenomenological models of many cortical information processing operations is
another strong argument for this theory.
Chapter 3, by Walter Freeman, is philosophically in tune with Chapter 1, but takes the main point even
further, invoking chaos as a key ingredient in brain dynamics. In Freeman's conception, each neuron
participates in multiple, interlinked, dynamical interactions with other neurons. Going farther than
Cauller, Freeman's theory views trajectories transitioning from one attractor to another as the
representations of information. The learning and behaving processes hinge on the occurrences of these
transitions, not on the sustained attainment of stable attractor limit cycles. Appropriate (goal-satisfying)
transitions are reinforced and enhanced, neutral or goal-denying transitions are suppressed. Freeman's
viewpoint has heft because its main tenets are behaviors that real neurons can probably achieve and
because it helps explain why conventional concepts have had such little success in explaining brain
information processing.
Chapter 4, by Robert Hecht-Nielsen, presents an almost entirely new, detailed and complete, high-level
theory of thalamocortical information processing. It postulates that all thalamocortical information
processing can be accounted for by two particular types of associative memory neural networks, each
employed in abundance. The first of these networks is a fast-acting two-step attractor system
(implemented by reciprocal connections between cortex and thalamus) which casts all information into
one of a few thousand fixed-for-life sparse neuronal codes (representational tokens). The cerebral
cortex is presumed to be tiled with tens of thousands of these (only weakly interacting) networks. The
second network learns to connect certain pairs of these tokens on different regions. The strength of
connection is a novel "fuzzy" information theoretic quantity called antecedent support; which Hecht-
Nielsen claims is the only information learned, stored, and used in the thalamocortex. The notion that
the cortex is fundamentally symbolic may appeal to old-school Al researchers. "Computational
Intelligence" fans may like the theory because the learned relationships between pairs of hard symbols
are soft and yet information-theoretic in character. Neuroscientists may like it because the whole thing
seems to exactly match the known anatomy and physiology of the thalamocortex and its related
structures.
Chapter 5, by Henry Markram, is a tour de force of advanced insight into those most critical brain
elements: synapses. Markram's synthesis of the huge and confusing literature on this subject is guided
by his own world-leading experimental findings. In particular, in strong resonance with several other
chapters of the book, he emphasizes the central importance of dynamics in synaptic transmission: the
post-synaptic effect of an arriving action potential depends critically on the recent history of arrivals of
other action potentials. Going far beyond a mere recitation of the facts, Markram presents a compelling
synthesis that identifies and characterizes principal categories of synaptic dynamical behavior. This
chapter provides abundant knowledge of great value to all interested in cortical theory.
Chapter 6, by Thomas McKenna, provides sage commentary on a variety of cortical theories and their
relationships. In particular, McKenna emphasizes and explores the gaps that clearly exist between
cortical theories and biology. These gaps are then used as pointers towards potentially profitable new
directions of theory construction. A valuable feature of McKenna's chapter is his characterizations and
comparisons of theories that operate at different levels of abstraction. Going far beyond the simple "top
down" and "bottom up" taxonomy, theories are discussed in terms of their ability to succeed at their
individual level of abstraction. Two hundred years from now any high school student will probably be
able to give a simple explanation for how the brain represents objects, learns and stores facts about
those objects and then uses those facts to carry out information processing. A key point is that a correct
theory can function without gaps at any level of detail. Another issue addressed in this chapter is the
applicability of cortical theories. For theories to be judged correct and complete they must support
practical implementations which can solve high-value human problems.
Chapter 7, by Jeffrey Sutton and Gary Strangman, proposes that the cerebral cortex can be viewed as
being composed of hierarchical groups of neural networks; or, as they term it, a Network of Networks
(NoN). The main point of this theory is that only modeling individual associative memory networks will
never, by itself, explain the function of the cerebral cortex. A key missing ingredient is that groups of
such networks are used, controlled, and trained as an ensemble—not individually. The chapter then
goes on to explain how this might be accomplished and why this view is so important in understanding
the obvious hierarchicality of processing we are so intimately familiar with (our cortex can deal with
"throw the ball" as a conceptual unit just as easily as "slightly increase bicep tension").
Chapter 8, by John Taylor, presents a theory of cerebral information processing that concentrates on
how consciousness may emerge from the combined activity of ensembles of stimulus-driven cortical-
subcortical control structures set up to tend to a variety of behavioral goals and needs. Taylor's analysis
frees itself from the details by fixating on firmly established high-level brain information properties that
any successful theory must explain. Mixing together insights from a variety of theoretical perspectives,
Taylor integrates these into a whole that emerges as a vision of how collections of relatively
straightforward "thermostat" brain functions can, because of the strong causal linkages between their
activating stimuli, end up supporting a higher-level process that becomes self-aware. Unlike past toy-
system-level studies of "emergent" properties, Taylor's theory gets to the heart of why and how
conscious awareness phylogenetically emerged. This chapter illustrates the value, and provides a
masterful example, of theory building at a high conceptual level.
Chapter 9, by Neill Taylor and John Taylor, addresses another key cortex theory issue: language. In
particular, they provide a foundational study of how a neuronal architecture can be developed to
implement Noam Chomsky's hypothetical "Language Acquisition Device." Central to their theory are
structures for chunking and unchunking of word strings (i.e., hierarchical networks). Using these, they
develop a set of neuronally plausible language processing components and procedures that exhibit
many of the key properties that a "LAD" must have. The resulting theory is not intended as a final theory
of language acquisition and understanding. However, it provides a nice jumping-off point for further
research.
Chapter 10, by Richard Zemel, is a theory of the population codes that many researchers believe lie at
the heart of cortical information processing. In the past, almost all studies of population codes have
assumed that the collection of neurons making up such a code are "ON" (and that other nearby neurons
are "OFF"). Even with this binary assumption, many interesting properties of such codes have been
established. In the theory presented in this chapter, Zemel goes way beyond this assumption: he allows
population codes to also have variable non-negative activity levels assigned to them. He then shows
that these activity levels can be interpreted in terms of a related probability distribution. Zemel then
argues that the process of activating a population code is itself a probability calculation. Furthermore,
he shows that collections of linked population codes can interact to form a type of belief network; a belief
network that seems highly compatible with cortical neuronal anatomy, physiology, and function. The net
result is a foundation for a new and promising direction of research.
Which theory do you prefer?
Chapter 1
Chapter 2
Chapter 3
Chapter 4
Chapter 5
Chapter 6
Chapter 7
Chapter 8
Chapter 9
Chapter 10
My own theory
 | |  | |  | |  |