Conference PaperPDF Available

Image Transformations in a Cognitive System. Tunnel transition and combining ensembles

Authors:

Abstract and Figures

The nature of creativity and its hidden mechanisms are areas that researchers have only recently begun to approach. In this paper, within the symbol-image cognitive architecture paradigm, we consider some operations on graphs that can be compared to the process of image synthesis and transformation in a cognitive system. We address such phenomena as the combination of neural ensembles and tunnel (sub-barrier) transition in a cognitive system. We also consider degree of fusibility of a cognitive system that characterizes its creative ability.
Content may be subject to copyright.
Image Transformations in a Cognitive System
Tunnel transition and combining ensembles
Ekaterina D. Kazimirova
Kaspersky Lab
Moscow, Russia
e-mail: Ekaterina.Kazimirova@kaspersky.com
Abstract The nature of creativity and its hidden
mechanisms are areas that researchers have only recently
begun to approach. In this paper, within the symbol-image
cognitive architecture paradigm, we consider some operations
on graphs that can be compared to the process of image
synthesis and transformation in a cognitive system. We address
such phenomena as the combination of neural ensembles and
tunnel (sub-barrier) transition in a cognitive system. We also
consider degree of fusibility of a cognitive system that
characterizes its creative ability.
Keywords: graph, symbol, attribute, cognitive architecture,
creativity.
I. INTRODUCTION
In the era of the rapidly developing ecosystem of living
things (the Internet of Things) and human-friendly
anthropomorphic robots [1][2][3], the issues related to
developing artificial intelligence systems that are not only
rational, but also creative, take on a special significance.
The ability of cognitive systems to generate new thoughts
and new images will turn machine into man's equal partner
in addressing important issues of the 21st century, in solving
its challenges, such as overcoming disease and resolving
complicated economic problems.
In this paper, we discuss some of the mechanisms that
could be employed in artificial cognitive systems to
generate new information. Such ability is treated as a
synonym for creativity.
The paper is organized as follows. A description of the
model is provided in Section II. An example of the model’s
application is provided in Section III. The summary and
conclusions are in Section IV.
II. DESCRIPTION OF THE MODEL
We consider the following organization of a cognitive
system. Each image encoded in the system is represented by
its attributes and the corresponding symbol. The attributes
are linked together by the image’s symbol [4], [5]. The main
role of symbolization is compressing information, but we
believe it is equally important that it prevents images with
the same attributes from being mixed together. This idea
was briefly described in [6].
For the sake of brevity, we consider a two-layer system
(graph), where the first layer encodes image attributes and
the second layer encodes image symbols.
A. Basic Relationships
In our model, the following relationships are formed
between a symbol and its attributes. When two or more
attributes are activated, the symbol is also activated.
Conversely, when a symbol is activated, the attributes of its
image are also activated.
Figure 1. The basic model represented by a two-layer graph. The markers
“Аand “Вcorrespond to symbols, 1, 2, 3, 4, 5 correspond to their
attributes.
In Figure 1, the basic two-layer model is presented. The
symbol "A" is activated when there are signals from
particular attributes (nodes 1, 2, or 3) in different
combinations, (provided that more than one node is
involved). The symbol "B" is activated when there are
signals from nodes 3, 4, or 5 in different combinations (also
provided that more than one attribute is involved). When the
symbol "A" is activated, this leads to the activation of an
image consisting of a set of attributes (1, 2, 3). Activation of
the symbol "B" leads to the activation of an image consisting
of a different set of attributes (3, 4, 5).
Let us consider the following example. The relations
between attributes and symbols in our model of a cognitive
system work similarly to children’s riddles. Something that
is round, striped, and sweet (nodes: attributes) is a
watermelon (node: symbol); something that is striped and
orange, with sharp claws and teeth, is a tiger.
Concerning the problem of the possibility of images
being mixed together, note that a cat and a dog have many
attributes in common. According to our model, it is symbols
that enable the cognitive system to tell them apart.
Let us illustrate the “more than one attribute” rule.
Given the attributes “striped” and “round”, we can guess
“watermelon” and given the attributes “striped” and “with
sharp teeth”, we can guess “tiger”. However, given only one
attribute (e.g., striped), we cannot guess what the object
is. And conversely, specifying a symbol (e.g., watermelon),
brings its attributes to mind (striped, round, sweet,
with seeds, etc.).
In a real-world system, the number of layers is obviously
much greater. For example, there is the integrative symbol
“carnivores” (Lat. Carnivora) above the symbols “cat” and
“dog”, etc.
B. Attribute and Attention
We assume that a cognitive process starts when one of
the cognitive system’s elements appears in the field of
attention. The development of a thought involves revealing
connections between that element and its neighbors, as well
as forming its new relationships and connections with other
elements of the cognitive system.
Figure 2. Activation of two symbols by one attribute.
According to our concept, prolonged activation of one
attribute (for a time interval t > tatt) increases its influence.
As a result, that attribute, even if it is alone, can activate the
associated symbols. In Figure 2, the attribute “3” is
activated by attention and, in turn, activates the symbols
“A” and “B”.
For example, thinking about speed, we can recall a car, a
cheetah, and an airplane.
Below we look at the ways in which a system based on
these rules can generate new information by transforming
images that is, creating new images out of existing ones.
C. Combining ensembles
Thus, the symbols associated with a certain attribute can
be activated by activating the attribute for a time interval t >
tatt. These symbols, in turn, activate the rest of their
attributes. As a result, a new ensemble combining the
elements of two images (symbols and attributes) is formed.
Figure 3. The activation of symbols A and B leads to the activation of all
their attributes. This gives rise to an ensemble ("A + B").
In Figure 3, the attribute 3 is enhanced by attention or
emotion, i.e., it receives additional activation from neurons
(represented by nodes in our model) that are external relative
to the ensembles A and B. Due to enduring activation, it
activates both neuron-symbols A and B simultaneously.
These two images, i.e., symbols "A" and "B" plus all their
attributes, are temporarily united into an ensemble. Starting
from that moment, the attributes of image B also belong to
the symbol A”, and vice versa (for a certain time interval).
This mechanism could serve as the basis for metaphorical
thinking (feature transfer). We briefly described this problem
in [7]. Issues related to the integration of information in the
cognitive system are also discussed in [8].
Let us consider the following example of metaphorical
thinking. In [9], the general metaphor argument is war” is
presented. G. Lakoff and M. Johnsen write, We see the
person we are arguing with as an opponent. We attack his
positions and we defend our own. We gain and lose ground.
We plan and use strategies. If we find a position
indefensible, we can abandon it and take a new line of
attack. What they describe in this passage is attribute
transfer between the concepts “war” and “argument”.
It is important to realize that the transfer process starts
after the common attribute of two concepts is found. In the
case of the metaphor “argument is war”, the term
confrontation represents the common attribute, while the
transfer process follows the mechanism described above.
G. Lakoff and M. Johnsen write, The most important
claim we have made so far is that metaphor is not just a
matter of language, that is, of mere words. We shall argue
that, on the contrary, human thought processes are largely
metaphorical.
D. Tunnel transition
Under certain conditions, the activation of attributes can
lead to transitions on the same (attribute) level (activation of
attributes through attributes) rather than the activation of
symbols. Suppose that there is a neurotransmitter acting at
the network’s attribute level, dynamically strengthening the
connections and thus facilitating the transition from one
attribute to another. Under such conditions, attributes
(neuron attributes) can activate each other without the
activation of symbols (see Figure 4). We called this effect
tunnel (sub-barrier) transition. In this case, the activation
wave can dive under an adjacent symbol and activate one
of the more remote symbols rather than the nearest one.
This process could be controlled by presence or absence
of a neurotransmitter. Favorable conditions for such a
process could also be created by the "constitutional"
characteristics of the cognitive system, e.g., relatively weak
connections between the attribute and symbolic levels
(attribute-symbol), or, on the contrary, by the relatively
strong connections at the attribute level (attribute-attribute).
Figure 4. Flow within the cognitive system.
We would like to emphasize that Hebb's rule [10] describes
the strengthening of connections between closely spaced
(directly contacting) neurons. In contrast, the tunnel
transition corresponds to the formation of connections
between non-neighboring neurons that are not directly
connected.
E. Fusibility of thinking
Within our model, the presence of a neurotransmitter in
a subnet containing attributes and/or the cognitive system’s
constitutional features make this system more "fusible", i.e.,
more fluid (akin to molten metal). It would be interesting to
study the "coefficient of fusibility (Kfus) of a cognitive
system defined as the ratio of the connection strength at the
attribute level to the strength of attribute-symbol
connections:
symbatrfus WWK /
(1)
Note that mental disorders can be associated with
different types of associative thinking impairment. For
example, the so-called acceleration of thinking (racing
thoughts) is characterized by an excessive emergence of
associations. As a result, thinking becomes superficial, with
attention being switched too easily.
Let us examine how the character of thinking depends
on Kfus. The creative process involves working with
associations as the mechanism of new image production.
We believe that the "coefficient of fusibility" may control
the ease with which associations emerge in the cognitive
system. At low values of Kfus, the "flow" (activation of
connections) between images on the same attribute level and
the subsequent merging of different images into new ones is
hampered, and the system becomes rigid. Such a system can
only work with the images (symbols together with their
attributes) it already contains. The system can analyze them,
i.e., it knows the properties of each symbol and can attribute
it. However, such a cognitive system is virtually incapable
of synthesizing new images. As Kfus increases, it becomes
possible for new images to appear within the cognitive
system. With Kfus > Kcritical, flow across the attribute level
becomes too easy, numerous associations arise, but they are
not fixed in new images. This process represents a thinking
disorder.
In psychology, our concept corresponds (to some extent)
to Raymond Cattell’s concept [11] of fluid and crystallized
intelligence, where crystallized intelligence is the ability to
operate with already acquired knowledge, skills and
experience, while fluid intelligence (or fluid reasoning)
corresponds to the ability to reason and solve novel
problems in new ways.
III. EXAMPLE
As an example, let us consider the poem The Soft
Moscow Rain by Osip Mandelstam, translated by Richard
McKane [12]:
It shares so stingily
its sparrow cold
a little for us, a little for the clumps of trees,
a little for the cherries for the hawker’s stall.
And a bubbling grows in the darkness,
the light fussing of tea-leaves,
as though an ant-hill in the air
were feasting in the dark green grass;
fresh drops stirred
like grapes in the grass,
as though the hot-bed of the cold
was revealed in web-footed Moscow.
Let us consider the main symbols and their important
attributes that are mentioned in this poem.
Figure 5. Semantic connections in Mandelstam’s poem The Soft Moscow
rain. "Tunnel transitions" result in unexpected combinations of symbols.
In Figure 5, we can see that symbols that are
semantically close to the term rain (such as cloud and
water”) do not appear in the poem. At the same time,
connections are established via attribute-attribute transitions
between semantically distant symbols, such as rain, “ant-
hill”, “tea leaves, and grapes. The artistic value of the
poem and the fact that it is a masterpiece seems to be due to
the “tunnel effect” which reveals distant connections, thus
helping to communicate an impression of the rain.
IV. CONCLUSION
Problems associated with understanding the mechanisms
of creativity and reproducing them form a barrier to creating
general Artificial Intelligence (AI), which has not been
overcome so far. It seems that entirely anthropomorphic AI
could only be developed if artificial intelligence systems
were able to think independently and perform creative tasks.
To achieve this, we will have to solve the problem of
generating new information in the cognitive system, which
was discussed within the framework of Dynamic Theory of
Information [13]. In this paper, we made an attempt to show
the possible basic mechanisms of information synthesis in
the cognitive system, illustrating them with some operations
on graphs. The main concepts discussed in this paper are:
formation of ensembles that combine different
images;
the "tunnel effect", i.e., the attribute-attribute
transition that leads to unexpected combinations of
symbols;
the "degree of fusibility" of a cognitive system.
The formation of ensembles and the “tunnel effect” are
associated with the mechanism that can transform images in
the cognitive system. In the former case (the formation of
ensembles), image transformation is caused by the transfer of
attributes from one image to another. In the latter case (the
“tunnel effect”), the attributes of semantically distant
symbols are combined together. This can result not only in
the enrichment of an existing image, but also in the
generation of a new image (e.g., as in Mandelstam’s poem,
where rain is associated with an ant-hill and grapes). Both
mechanisms are characteristic of cognitive processes, not the
simple image classification provided by existing artificial
neural networks. It should be emphasized that in both of the
above cases, new information is generated. These
mechanisms could be closely connected with the intuitive
and creative thinking process.
The concept of the degree of fusibility, as well as the
coefficient of fusibility introduced in this paper, when
applied to an artificial cognitive system, could provide a way
of controlling the rigidity of artificial cognitive systems,
making them more adept at reflecting the reality or, on the
contrary, more intuitive and creative.
Implementing these mechanisms could help to achieve an
AI that can think creatively and has intuition. The next step
would be to develop these ideas further as a mathematical
model.
ACKNOWLEDGMENTS
The author is grateful to Olga Chernavskaya, Evgeny
Volovich and Artem Vorontsov for the fruitful discussions.
REFERENCES
[1] What is Amazon Alexa? Available from:
https://developer.amazon.com/alexa Retrieved 2018.01.09
[2] Meet your Google Assistant. Available from:
https://assistant.google.com. Retrieved 2018.01.09
[3] Azuma Hikari. Official Site. Available from:
http://gatebox.ai/hikari/en/ Retrieved 2018.01.09
[4] O. D. Chernavskaya, D. S. Chernavskii, V. P. Karp, A. P.
Nikitin, and D. S. Shchepetov, “An architecture of thinking
system within the Dynamical Theory of Information,” BICA,
vol. 6, pp. 147158, 2013.
[5] O. D. Chernavskaya, D. S. Chernavskii, V. P. Karp, Ya. A.
Rozhylo “An architecture of the cognitive system with
account for emotional component,” Biologically Inspired
Cognitive Architectures, vol.12, pp. 144154, 2015.
[6] E. D. Kazimirova, “Two-Component Scheme of Cognitive
System Organization: the Hippocampus Inspired Model,” The
Ninth International Conference on Advanced Cognitive
Technologies and Applications (COGNITIVE 2017) IARIA,
Feb. 2017, pp. 2123, ISBN: 978-1-61208-531-9
[7] E. D. Kazimirova, “Elements of the symbol-image
architecture of cognition and their parallelism to certain
linguistic phenomena”, Neurocomputers, vol. 4, pp. 3537,
2015. Available from: http://www.radiotec.ru/article/16364
Retrieved 2017.12.12
[8] M. Oizumi, L. Albantakis, and G. Tononi, “From the
Phenomenology to the Mechanisms of Consciousness”
Integrated Information Theory. Available from:
http://journals.plos.org/ploscompbiol/article?id=10.1371/jour
nal.pcbi.1003588 Retrieved 2018.01.09.
[9] G. Lakoff and M. Johnsen, “Metaphors we live by”, London:
The university of Chicago Press, pp. 810, 2003.
Available from: http://shu.bg/tadmin/upload/storage/161.pdf
Retrieved 2018.01.24
[10] D. O. Hebb. The organization of behavior: a neuro-
psychological theory. New York, 2002.
[11] R. B. Cattell, "Theory of fluid and crystallized intelligence: A
critical experiment", Journal of Educational Psychology,
vol. 54, pp. 122, 1963.DOI:10.1037/h0046743
[12] Available from: http://arlindo-correia.com/200801.html
Retrieved 29.01.2018
[13] D. S. Chernavsky, Synergetics and information: Dynamic
Theory of Information. Moscow: URSS, 2004. (in Russian).
(Sinergetica i Informaciya: Dinamicheskaya Theoriya
Informacii)
ResearchGate has not been able to resolve any citations for this publication.
Article
This article represents an extension of authors’ previous papers (Chernavskaya, Chernavskii, Karp, Nikitin, & Shchepetov, 2012, 2013) in modeling cognitive systems on the base of the Dynamical Theory of Information. The paper focuses on the problem of account for emotions in artificial system. The main hypothesis consists in the assumption that emotions inherent in a living system could be simulated by variation of amplitude of the occasional component (noise) inherently embedded into the architecture of artificial system. Within this concept, increasing noise amplitude should correspond to negative emotions (anxiety), while its decreasing provides positive emotions (relaxation, pleasure). A rapid up-and-down spike in the noise amplitude could imitate a laugh. This hypothesis is secured by incorporation of an additional dynamical variable that represents an analogy to the compound of neural mediators in human beings. The system of linked equations in terms of “noise amplitude – neurotransmitter compound” is proposed to describe mutual influence of the cognitive process and emotional component. The model permits to reproduce qualitatively certain prominent effects typical for human emotional reactions (like stress and shock).
Article
The problem of the thinking process modeling is considered within the context of Dynamical Theory of Information. The proposed definition for thinking process is based on listing its main goals and functions. A set of various type neural processors is considered to reveal necessary and sufficient combination to implement these basic functions. It is shown that, according to DTI, the artificial thinking system is to be divided into two subsystems, where the first subsystem provides generation of information and learning, the second one—storage and application of the information. This correlates with presence of two hemispheres of the human brain, one of them is believed to produce the intuition, while the other one being responsible for the logical thinking. A particular version of the thinking-system architecture is proposed. It is shown that in the course of evolution, the system is progressively growing from low imaginary information levels, through semantic levels (where the information has sense for this individual system only), up to high abstract levels that could be verbalized to be generally realized. An interpretation of intuitive and logical thinking is given. Our approach, being close to that common for BICA, does differ in certain points. We are not dealing with the active agents that are a priori endowed with emotions, motivations, and comprehension. Instead, starting with just formal (but dynamical) neurons, the system itself, due to the self-organizing nonlinear interactions of the neuron ensembles, could produce semantic and abstract information, and perceive the verbalized one.
Article
The theory of fluid and crystallized intelligence propounded 20 years ago by the author has since accumulated support. However, the crucial issue of whether 1 or 2 general factors subtend intellectual performances has lacked an experiment adequately designed for accurate, determinate, simple-structure rotation at the 2nd order. By factoring culturally embedded with culture-fair intelligence measures on a background of pure personality primaries (N = 277 7th and 8th grade boys and girls), it is shown that 2 general factors indeed exist. A review, with some mathematical formulations, is given of the theory's implications for the nature-nurture ratio, brain injury, standard deviaiton of the IQ, growth curves, the concept of a relational difficulty hierarchy, test standardization, and the relative validities of traditional and culture-fair intelligence tests. (PsycINFO Database Record (c) 2012 APA, all rights reserved)
Two-Component Scheme of Cognitive System Organization: the Hippocampus Inspired Model
  • E D Kazimirova
E. D. Kazimirova, "Two-Component Scheme of Cognitive System Organization: the Hippocampus Inspired Model," The Ninth International Conference on Advanced Cognitive Technologies and Applications (COGNITIVE 2017) IARIA, Feb. 2017, pp. 2123, ISBN: 978-1-61208-531-9
Elements of the symbol-image architecture of cognition and their parallelism to certain linguistic phenomena
  • E D Kazimirova
E. D. Kazimirova, "Elements of the symbol-image architecture of cognition and their parallelism to certain linguistic phenomena", Neurocomputers, vol. 4, pp. 3537, 2015. Available from: http://www.radiotec.ru/article/16364 Retrieved 2017.12.12
From the Phenomenology to the Mechanisms of Consciousness
  • M Oizumi
  • L Albantakis
  • G Tononi
M. Oizumi, L. Albantakis, and G. Tononi, "From the Phenomenology to the Mechanisms of Consciousness" Integrated Information Theory". Available from: http://journals.plos.org/ploscompbiol/article?id=10.1371/jour nal.pcbi.1003588 Retrieved 2018.01.09.
Metaphors we live by
  • G Lakoff
  • M Johnsen
G. Lakoff and M. Johnsen, "Metaphors we live by", London: The university of Chicago Press, pp. 8-10, 2003. Available from: http://shu.bg/tadmin/upload/storage/161.pdf Retrieved 2018.01.24
Synergetics and information: Dynamic Theory of Information
  • D S Chernavsky
D. S. Chernavsky, Synergetics and information: Dynamic Theory of Information. Moscow: URSS, 2004. (in Russian). (Sinergetica i Informaciya: Dinamicheskaya Theoriya Informacii)