Challenging the Intuition about Memory and Computation in Theories of
Cognition
Jochen Kerdels and Gabriele Peters
FernUniversit
¨
at in Hagen, University of Hagen, Human-Computer Interaction,
Faculty of Mathematics and Computer Science, Universit
¨
atsstrasse 1, D-58097 Hagen, Germany
Keywords:
Memory, Computation, Theories of Cognition, Cognitive Architectures, Neurobiology.
Abstract:
In this position paper we argue that the concepts of memory and computation as they are commonly used
in theories of cognition are strongly influenced by our intuitive understanding of the corresponding concepts
in contemporary computer systems, leading to an implicit loss of biological plausibility. To support our ar-
gument we provide an alternative perspective on memory and computation that allows a closer comparison
of the capabilities of computer programs running on computer systems and neurobiological systems showing
that computer programs exhibit a computational flexibility that is difficult to reconcile with neurobiological
constraints. We end this paper by offering a neurobiologically plausible perspective on memory that views
memory as a dynamic, distributed process that is an intrinsic part of a neurobiological network that integrates
information, e.g., sensory information.
1 INTRODUCTION
Theories of cognition seek to present models of the
human mind that enlighten our understanding of a
wide spectrum of cognitive processes including inte-
gration of sensory information, generation of behav-
ior, and processes of high-level thought. Cognitive
architectures in particular aim at providing compu-
tational instantiations of such models that facilitate
the exploration of dynamic properties and behaviors.
Prominent examples of such architectures are ACT-
R (Anderson, 1996), SOAR (Laird et al., 1987; Laird,
2008), CLARION (Sun, 2007), LIDA (Franklin et al.,
2013), or the highly influential Global Workspace
Theory (GWT) by Bernard Baars (Baars, 2017).
In this paper we would like to discuss two basic
assumptions that commonly underlie theories of cog-
nition that seem to be true almost trivially but may
impact the neurobiological plausibility of these theo-
ries severely. The first assumption concerns the con-
cept of memory, the second assumption concerns the
related concept of computation. We will argue that
the concept of memory as it is commonly used in the-
ories of cognition is strongly influenced by our un-
derstanding of memory as it is used in typical com-
puter systems with von Neumann architecture. It is
assumed that memory is a kind of container that al-
lows to store pieces of information and to retrieve
these pieces again later on. Similarly, the second as-
sumption that we would like to challenge concerns the
concept of computation. It is the idea that pieces of
information present in such a memory can be used
by different parts of a cognitive system implying an
encoding of information whose meaning is shared by
these different parts. We hope to show that both as-
sumptions, when scrutinized, are difficult to reconcile
with the properties and constraints inherent to a plau-
sible neurobiological substrate. As an alternative, we
present a different perspective that views memory as
a process that is an intrinsic, distributed part of a neu-
robiological system.
The following section highlights typical ideas of
memory and its usage in a number of cognitive ar-
chitectures, and relates these ideas to a common and
widespread intuition about computation. Section 3
elaborates on this intuition and proposes an alterna-
tive perspective on computation that views computer
programs as transformation networks, which enable
a closer comparison of the computational capabili-
ties of general purpose computer systems and neu-
robiological systems. Section 4 then proposes an
alternative perspective on the structure and function
of memory that emphasizes its integrative and dis-
tributed characteristics. Finally, section 5 concludes
this position paper.
522
Kerdels, J. and Peters, G.
Challenging the Intuition about Memory and Computation in Theories of Cognition.
DOI: 10.5220/0008493605220527
In Proceedings of the 11th International Joint Conference on Computational Intelligence (IJCCI 2019), pages 522-527
ISBN: 978-989-758-384-1
Copyright
c
2019 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
2 MEMORY IN THEORIES OF
COGNITION
Memory is a central part of many theories of cogni-
tion where it is used for multiple purposes. Amongst
other uses, it can serve as a communication hub that
allows to dynamically route information between dif-
ferent parts of the cognitive system and establish a
kind of global context (seen as analog to “short term
memory”), it can serve as a storage of past expe-
riences (“episodic memory”) and knowledge about
the world (“declarative memory”), or it can serve as
a mechanism that facilitates the acquisition and re-
production of behavioral patterns (“procedural mem-
ory”). In all these cases, either explicitly or implicitly,
memory is typically seen as a container that can store
and reproduce signal patterns much like the memory
in a computer system that stores and reproduces bit
patterns. In particular, pieces of information are seen
as self-contained entities that can “move around” and
exist at different locations within the computational
system. This notion is reflected by, e.g., John Laird
(SOAR) as he outlines what a cognitive architecture
consists of in his view (Laird, 2008):
A cognitive architecture consists of:
memories for storing knowledge
processing units that extract, select, com-
bine, and store knowledge
languages for representing the knowledge
that is stored and processed
[..] Cognitive architectures must embody
strong hypotheses about the building blocks
of cognition that are shared by all tasks, and
how different types of knowledge are learned,
encoded, and used, making a cognitive archi-
tecture a software implementation of a general
theory of intelligence.
Another example that highlights this perspective is
given by John Anderson (ACT-R) in the context of
describing how knowledge about the environment is
processed in their system (Anderson, 1996):
We have only developed our ideas about en-
vironmental encodings of knowledge with re-
spect to the visual modality. In this area,
it is assumed that the perceptual system has
parsed the visual array into objects and has
associated a set of features with each object.
ACT-R can move its attention over the visual
array and recognize objects. [..] Features
within the spotlight can be synthesized into
recognized objects. Once synthesized, the ob-
jects are then available as chunks in ACT’s
working memory for further processing.
Finally, a good example of how memory is utilized
as a central mechanism to coordinate and control the
flow of information in a cognitive system is given by
Bernard Baars as he describes the core idea of his
Global Workspace Theory (Baars, 2017):
Global Workspace Theory (GWT) suggests
that the brain has a fleeting integrative ca-
pacity that enables access between functions
that are otherwise separate. This makes sense
in a brain that is viewed as a massive paral-
lel set of highly specialized neuronal proces-
sors. In such a system coordination and con-
trol may take place by way of such a central
information exchange, allowing some special-
ized processors such as sensory regions in
cortex to distribute information to the system
as a whole. This solution also works in large-
scale computer architectures, which show typ-
ical “limited capacity” behavior when infor-
mation flows by way of a global workspace.
These quotes illustrate how the concept of memory in
cognitive architectures is closely aligned with the cor-
responding concept in contemporary computer sys-
tems. Linked to this understanding of memory is a
matching intuition about computation. In its most ba-
sic form this intuition about computation consists in
the idea of executing a list of basic operations that
each take some data as input and produce some data
as output that affects the subsequent steps of the com-
putation in a deterministic way, e.g., by generating
new input data, controlling the program flow, or caus-
ing desired side effects. For many, this intuition about
computation matches their introspective understand-
ing of the apparent sequential nature of high-level,
linguistic thought or the planning and control of de-
liberate actions necessary to perform a complex task.
One important aspect of this analogy between com-
putation and high-level mental processes is the sense
of agency that the latter imply. We think thoughts.
We perform actions. The computer
1
processes data.
This implicitly assumed existence of agency makes
thoughts and data alike into objects that are acted
upon. This object quality in turn implies the existence
of such objects as coherent entities in some medium.
In case of data this medium appears to be memory.
This basic intuition about computation is a pow-
erful tool, an abstraction, to reason about and write
complex computer programs, and it is tempting to use
it to interpret and model the function of neurobiolog-
ical information processing systems as well. How-
ever, we argue that the latter use of this abstraction is
counterproductive and leads to misleading analogies
1
or its central processing unit (CPU)
Challenging the Intuition about Memory and Computation in Theories of Cognition
523
between computer and brain, especially regarding the
purpose and function of memory in both systems. To
elucidate this aspect the next section provides an alter-
native intuition about memory and computation that
allows for a more direct comparison of computer and
neurobiological systems. It will highlight fundamen-
tal differences of both systems regarding their respec-
tive capabilities to process information.
3 COMPUTER PROGRAMS AS
TRANSFORMATION
NETWORKS
The intuition about computation outlined above is
supported and reinforced by modern programming
languages as they facilitate the definition of com-
plex data types and their instantiation as objects that
seem to exist as such entities in memory. In addi-
tion, these objects are commonly described in a way
that make it seem as if these objects exhibit forms
of agency by possessing a local, encapsulated state,
and being responsible for a specific set of subprob-
lems. This useful abstraction allows a programmer to
describe the solution to a computational problem in
a way that is both meaningful to other programmers
and herself as well as “understandable” by the com-
puter system. This “understanding”, however, is for
the most part just an illusion. When a program is ac-
tually loaded into memory and executed, the resulting
bit patterns that are stored in memory are by them-
selves meaningless. Their meaning is only encoded
in the program that reads and interprets these patterns,
and much of this encoded meaning is already stripped
away when the high-level program is translated into
machine code. More precisely, most of the meaning
attributed to certain high-level program structures by
a programmer are not part of the program that will ac-
tually run on the computer. In that sense, the “actual
meaning” of a bit pattern in memory is as elusive to
the computer system as an evolved, instinctive behav-
ior might be to an insect. Yet, the implicit sense of
agency mentioned above is suggestive of locating the
meaning of a bit pattern intuitively somewhere within
the machine rather than within the programmer who
is describing and interpreting the machine’s behavior.
To gain an alternative perspective that avoids this
intuitive, but in our view misleading interpreta-
tion of a program we take a closer look at the idea of
computation as executing a list of basic operations. In
essence, such a basic operation can be understood as a
prescription of how a particular input pattern is trans-
formed into a new output pattern given a limited set
of hardware resources. Using such basic operations
over and over again to build a complex web of input
output transformations is what a computer program
does. The “glue” that enables the computer program
to form this network of transformations is memory.
When a computer program executes it orchestrates
a constant circulation of bit patterns between memory
and CPU. By doing so the limited hardware resources
of the CPU are reused over and over again to suc-
cessively perform all the input output transformations
in the network of transformations that is defined by
the program. Hence, every computer program can be
seen as the emulation of a vastly more complex, spe-
cial purpose CPU that would implement every input
output transformation defined in the program in hard-
ware. In that sense, every memory retrieval within a
computer program - even retrieval from distant mem-
ories like a hard disk or some network resource - cor-
responds to a dynamic reconfiguration and/or expan-
sion of this virtual, special purpose CPU.
In this alternative view on computation memory
serves as a pattern buffer between transformations
that endows the resulting, virtual transformation net-
work with a number of capabilities that would be
challenging or impossible to implement in a real,
physical network in which the hardware performing
each transformation is connected directly. In particu-
lar, the use of memory enables the network to oper-
ate independent of timing constraints, i.e., the output
of a transformation can be buffered for an arbitrary
amount of time before it is processed by further trans-
formations. Typical uses of this capability include
waiting for the results of multiple transformations
before proceeding, combining the output of a trans-
formation with its past outputs, or even distributing
a computation over multiple devices. Furthermore,
memory enables the virtual transformation network to
have arbitrary connectivity patterns that are not bound
to physical constraints like limited signal speed, lim-
ited fan-out and fan-in of the transformation hard-
ware, or limited physical space. Together, these ca-
pabilities of the virtual transformation network reflect
the flexibility and power of computer programs run-
ning on contemporary computer systems. The key
to this flexibility and power is the use of memory as
the central interface mechanism for this transforma-
tion network. We argue that this view on computa-
tion provides a more accurate and simpler picture of
the function and purpose of memory than the high-
level perspective provided by modern programming
languages where the apparent program structure dif-
fuses into the structure of the virtual transformation
network when the program is translated into machine
code.
NCTA 2019 - 11th International Conference on Neural Computation Theory and Applications
524
Using this alternative view on computation allows
a more direct comparison to the processing of infor-
mation in neurobiological systems. From an evo-
lutionary perspective neurobiological systems devel-
oped to solve specific computational problems, e.g.,
to control muscles or to process the information com-
ing from a sensory system. As such, the respec-
tive networks of neurons or groups of neurons could
be seen as specialised transformation networks sim-
ilar to the transformation networks described above.
However, instead of being emulated and virtual, these
natural transformation networks are actually imple-
mented within the neurobiological substrate removing
the necessity for memory as a “transformation glue”
while at the same time adding a number of constraints
regarding possible connection patterns and types of
transformations. In particular, an on-the-fly reconfig-
uration and expansion of the neurobiological transfor-
mation network is not possible in general. Instead, the
basic structure of the network emerges during the de-
velopment of the organism and is controlled by, e.g.,
periods of cell proliferation, cell migration along cel-
lular support structures, or guidance by short and long
range chemical gradients (Squire et al., 2008). The
variability inherent in this self-organized formation
limits the extent with which the solution to a specific
computational problem, e.g., a behavioral pattern, can
be hardwired into the network structure and adapted
over time. To mitigate this constraint neurobiologi-
cal networks show forms of plasticity that facilitate
fast changes to the network by adapting the response
of individual neurons or local groups of neurons to a
given input signal. As a consequence, the correspond-
ing input output transformation implemented by those
neurons changes accordingly. However, subsequent
transformations are not informed about this change
but have to adapt themselves as well if necessary and
propagate the change further. This means that from a
global perspective the way by which signals are en-
coded and processed within the transformation net-
work is only known locally and remains in a constant
flux.
This dynamic change of how information is en-
coded and processed by different parts of the trans-
formation network conflicts with an idea of memory
that sees memory as a container that stores patterns
of information for later use since the encoding of that
information might have changed while it was stored.
Such an outdated encoding would then become unin-
telligible for the network. Similarly, the idea of mem-
ory as a means to coordinate and control the flow of
information relies as well on a consistent encoding,
which is not guaranteed when local neuroplasticity is
taken into account. Despite these doubts regarding
the biological plausibility of a container-like memory
one might argue that neurobiological systems clearly
do have the ability to remember, e.g., past experiences
and therefore must have some form of memory. To
address this point we will outline our view on a neu-
robiologically plausible memory in the next section.
4 MEMORY AS A PROCESS
In the previous section we described how local neuro-
plasticity continually changes the way how neurons
or local groups of neurons respond to their inputs
and thus encode these inputs differently for neurons
downstream in the network. A memory system that is
based on storing and recreating patterns of informa-
tion is not well suited to cope with this drift of encod-
ing. We therefore suggest that memory in a neurobio-
logical network does not store and recreate the signals
that pass through the network but is rather a mecha-
nism that allows to store and reestablish the activation
state of the network or parts thereof, i.e., to not recre-
ate the result of some computation but to reestablish
the conditions that led to the result.
We argue that such a type of memory has to be a
distributed, intrinsic part of the network instead of be-
ing a dedicated, localized memory system. Moreover,
we see local neuroplasticity, the characteristic respon-
sible for the drift of encoding, as a key mechanism for
this memory. It enables individual neurons to learn
typical input patterns that capture some information
about the statistical nature of their inputs (Kerdels and
Peters, 2018). If such a neuron n
a
receives its inputs
from sensory cells, then the neuron learns something
about the statistical nature of that part of the world
that is transduced by these cells. If the neuron n
a
receives its inputs from multiple other neurons n
i
, it
learns something about the statistical nature of the co-
activation of these input neurons. It learns or forms
an association between these cells, i.e., it becomes a
small associative memory. However, at this stage it
is more of an association detector or something akin
to a hash function. The cell could answer the ques-
tion “Have I seen this activation pattern before?”, but
it can not be “read out” to reestablish that activation
pattern. A solution to this problem arises when one
assumes that the neurons n
i
are capable of forming as-
sociative memories themselves. In that case, the orig-
inal associative memory neuron n
a
would just have
to form reciprocal feedback connections to all its in-
puts. Reading out this cell, i.e., activating it, would
then result in reestablishing the activation state of the
input neurons n
i
. Although this reciprocal connec-
tion appears rather simple it can exhibit a range of
Challenging the Intuition about Memory and Computation in Theories of Cognition
525
interesting behaviors in relation to memory function
depending on the balance of activation between the
cell n
a
and its inputs n
i
. If the feedback connection
between n
a
and n
i
is inactive then signals can flow
in a bottom up direction allowing the neuron n
a
to
learn typical activation patterns of the neurons n
i
. If
the feedback connection becomes active the neuron n
a
can keep the neurons n
i
in a specific pattern of activa-
tion even if the original input signals to the neurons n
i
are no longer present, i.e., the cells n
a
and n
i
behave
now like a small short term memory of the original
inputs to the cells n
i
. Lastly, if the feedback of neu-
ron n
a
is dominant it can force the cells n
i
to take on a
pattern of activation that might disagree with the cur-
rent bottom up input to the neurons n
i
. In this case the
behavior resembles the recall of a long term memory
that overrides the current, bottom up driven state.
Viewed in isolation it might appear as a bold claim
that the small neuronal circuit described above would
constitute a proper memory mechanism. In particular,
the capacity of the single cell n
a
to learn and represent
the activity patterns of its inputs n
i
seems too limited.
This issue can be addressed by expanding this basic
mechanism to a local group of cells g
a
that exhibit a
form of competitive Hebbian learning via local inhi-
bition. It can be shown that the resulting ensemble
activity of such a group has a significantly higher ca-
pacity to learn and represent the activity patterns of
their inputs (Kerdels and Peters, 2017). Expanding
this model further one might imagine a network of
such reciprocally connected groups of neurons that
interact with each other according to the dynamics
described above. Depending on the respective influ-
ences between the groups such a network would be
continually driven towards states of temporarily sta-
ble, local equilibria that integrate current bottom up
signals from, e.g., sensory cells, hold onto parts of
this bottom up stream at appropriate stages of integra-
tion using local short term memory, and augment this
integrated information by forcing activity patterns top
down that correspond to a form of long term knowl-
edge. Reaching a temporary equilibrium in this net-
work could signal an episodic memory mechanism
like the hippocampus in mammals to capture the ac-
tivity of a small subset s of neuron groups at key lo-
cations within the network. When the activity pat-
tern of this subset s is reestablished again, it would
then force large parts of the network into a configura-
tion that would approximate the state of the network
when the episodic memory was formed. In principle,
the episodic memory itself could rely on similar basic
mechanisms of associative pattern recognition as de-
scribed above.
In summary, we argue that memory in a neurobio-
logical network is a distributed, dynamic process that
is an intrinsic part of the network both structurally and
functionally. This perspective of memory is in strong
contrast to the concept of memory prevalent in well-
known cognitive architectures that see memory as a
distinct module that stores and retrieves data like the
memory in a contemporary computer system. Inte-
grating neurobiologically plausible forms of memory
and computation in theories of cognition is a chal-
lenging task as it requires a significant shift in the
perspective on typical computational paradigms. A
recent example of a first step at such an integration is
the work of Hamid and Braun who suggest to infuse
the theory of reinforcement learning with the frame-
work of attractor neural networks in order to benefit
from the latter’s capacity of pattern-completion and
noise-insensitivity (Hamid and Braun, 2019).
5 CONCLUSION
In this position paper we argued that common intu-
itions about memory and computation adversely af-
fect how we think about the processing of information
in neurobiological systems, and how this perspective
is reflected by prominent cognitive architectures. We
provided an alternative view on memory and compu-
tation to show more clearly how these intuitions imply
computational capabilities that are difficult to recon-
cile with the constraints inherent in neurobiological
systems. In order to present a perspective on memory
that is more neurobiologically plausible we outlined
the idea of memory as an intrinsic, distributed part of
a neurobiological network whose dynamic processes
combine aspects of memory and computation in a
way that questions the otherwise sharp delineations of
these two concepts present in contemporary theories
of cognition.
REFERENCES
Anderson, J. R. (1996). Act: A simple theory of complex
cognition. American Psychologist, (51):355–365.
Baars, B. (2017). The Global Workspace Theory of Con-
sciousness: Predictions and Results, pages 227–242.
Franklin, S., Madl, T., D’Mello, S., and Snaider, J. (2013).
Lida: A systems-level architecture for cognition, emo-
tion, and learning. IEEE Transactions on Autonomous
Mental Development, 6.
Hamid, O. H. and Braun, J. (2019). Reinforcement Learn-
ing and Attractor Neural Network Models of Associa-
tive Learning, pages 327–349. Springer International
Publishing, Cham.
NCTA 2019 - 11th International Conference on Neural Computation Theory and Applications
526
Kerdels, J. and Peters, G. (2017). Entorhinal grid cells may
facilitate pattern separation in the hippocampus. In
Proceedings of the 9th International Joint Conference
on Computational Intelligence, IJCCI 2017, Funchal,
Madeira, Portugal, November 1-3, 2017., pages 141–
148.
Kerdels, J. and Peters, G. (2018). A grid cell inspired
model of cortical column function. In Proceedings
of the 10th International Joint Conference on Com-
putational Intelligence (IJCCI 2018), Seville, Spain,
September 18-20, pages 204–210.
Laird, J. E. (2008). Extending the soar cognitive architec-
ture. In Proceedings of the 2008 Conference on Arti-
ficial General Intelligence 2008: Proceedings of the
First AGI Conference, pages 224–235, Amsterdam,
The Netherlands, The Netherlands. IOS Press.
Laird, J. E., Newell, A., and Rosenbloom, P. S. (1987).
Soar: An architecture for general intelligence. Arti-
ficial Intelligence, 33(1):1 – 64.
Squire, L., Bloom, F., Spitzer, N., Squire, L., Berg, D.,
du Lac, S., and Ghosh, A. (2008). Fundamental Neu-
roscience. Fundamental Neuroscience Series. Elsevier
Science.
Sun, R. (2007). The importance of cognitive architectures:
an analysis based on clarion. Journal of Experimental
& Theoretical Artificial Intelligence, 19(2):159–193.
Challenging the Intuition about Memory and Computation in Theories of Cognition
527