Annals of Mathematics and Artificial Intelligence 26 (1999) 215252

215

Perceive this as that Analogies, artificial perception, and

category theory
Zippora Arzi-Gonczarowski
Typographics Ltd, 46 Hehalutz st, Jerusalem 96222, Israel
E-mail: [email protected]

This paper formalizes and analyzes cognitive transitions between artificial perceptions
that consist of an analogical or metaphorical transference of perception. The formalization
is performed within a mathematical framework that has been used before to formalize other
aspects of artificial perception and cognition. The mathematical infrastructure consists of
a basic category of artificial perceptions. Each perception consists of a set of world
elements, a set of connotations, and a three valued (true, false, undefined) predicative
connection between the two sets. Perception morphisms describe structure preserving
paths between perceptions. Quite a few artificial cognitive processes can be viewed and
formalized as perception morphisms or as other categorical constructs. We show here
how analogical transitions can be formalized in a similar way. A factorization of every
analogical transition is shown to formalize metaphorical perceptions that are inspired by the
analogy. It is further shown how structural aspects of better analogies and metaphors can be
captured and evaluated by the same categorical setting, as well as generalizations that emerge
from analogies. The results of this study are then embedded in the existing mathematical
formalization of other artificial cognitive processes within the same premises. A fallout of
the rigorous unified mathematical theory is that structured analogies and metaphors share
common formal aspects with other perceptually acute cognitive processes.

1.

Introduction

Cognitive transitions between settings that are perceived as analogical, in some

sense, constitute a salient natural activity of human intelligence. They have been
recognized and studied since antiquity. In an AI context the fruitfulness of analogies should typically depend on whether any testable consequences could be deduced
from them. Analogizing could sometimes be useful as a cognitive tool for perceptive
artificial intelligent agents.
Analogies may explicate unfamiliar environments and situations in terms of more
familiar ones. For example, an intelligent artifact that has been trained to perceive
a bookstore environment and interact with (e.g., shop in) this environment, can be
efficiently taught to do the same in a music store, with the necessary analogs being
drawn. Likewise, an intelligent artifact that has been trained to perceive music and
interact with (e.g., play on) the keyboard of a piano can be efficiently taught to
do the same with another keyboard instrument, the necessary analogs being drawn.
J.C. Baltzer AG, Science Publishers

216

Z. Arzi-Gonczarowski / Perceive this as that Analogies

In rational thinking, analogies may be used to suggest hypotheses about a more

general rule or setting. For example, an analogy between a bookstore environment
and a music store environment may suggest to an observant intelligent artifact general patterns of behavior of agents selling and/or shopping in store environments.
Consequences could be deduced for general store environments. Likewise, the analogy between a piano and another keyboard instrument may suggest to an observing
intelligence general patterns of music playing. Consequences could be deduced for
keyboard instruments in general.
Linguistic translations constitute examples of analogies. One may look at a language
as an environment that consists of perceptible audio/visual elements (words, phrases,
etc.). A translation is thus a cognitive transition to the respective environment of
another language. Analogs are typically being drawn between linguistic elements
that have the same semantic properties, but they can also preserve (or consistently
map) the audio properties of the linguistic element, (such as measure and rhyme in
translations of poetry), cultural insinuations [51], or even visual properties of the
linguistic element, if there is, for example, a perceptual sensitive transference of
font.
In their interaction with human agents, intelligent artifacts might need to analogize
when they follow a human line of reasoning (e.g., in the process of expert knowledge
acquisition). Metonymies, metaphors, tropes and a variety of other expressive tools
that are based on analogies have pervaded human cognitive processes to a point
that it would be hard to avoid them when communicating with artificial systems.
Lakoff [3941], for example, argues that language is tropological, and that its usage
is typically inspired by analogy to bodily experiences, as in arriving at a solution
or seeing a point.
A mathematical theory of artificial perceptions is proposed as a framework for the
formalization of cognitive transitions between settings that are analogical in some
perceptible sense. It is shown that the theory can naturally formalize analogy-making,
often a useful cognitive activity.
The body of research about analogies in thought, science, language, and other
cognitive domains and activities is so vast that it is hard to say anything new, that
has not already been said before, about analogies. This paper neither competes with
existing theories (some of them will be briefly sketched in the next section), nor does it
take sides in debates among various philosophical stances (e.g., traditional comparison
theories, interactionist theories, and others). The force of a categorical approach is
in avoiding over determination [48], and hence the paper will also not deal with
various semantic distinctions between analogies, similes, models and other related
phenomena. The novelty and the goal of the proposed approach is the introduction
of a rigorous mathematical categorical framework, where no such framework already
exists. In addition to meticulous mathematical rigor, this approach opens the way for
the integration of analogy making with other cognitive processes that are based on the
same mathematical formalism.

Z. Arzi-Gonczarowski / Perceive this as that Analogies

2.

217

Background and related works

2.1. Models of analogy-making

Analogy making has always been a natural human activity, and biblical parables
(typically euphemistic in nature) constitute early examples. Likewise, the recognition
and the study of this multifaceted phenomenon date back to antiquity. Philosophers
and scientists have continuously employed and studied this versatile cognitive process.
Artificial intelligence and cognitive science have lately aroused a renewed interest in
analogies from their perspectives. Thagard [55] counts the study of analogies as one
of the fundamental theoretical approaches to mental representation.
Making and following analogies in an artificial system necessitates a meticulous
formulation of greater rigor than the verbal descriptions that had been employed before
AI. One is now faced not only with the goal of recognition and critical analysis of
analogies and metaphors, but also with a new potential task of the generation and
creative synthesis of analogies and metaphors by an artificial intelligent system. One
of the first attempts in this direction was Thomas Evans program ANALOGY, described
in [49], that was designed to solve visual geometric analogy problems in which it had
to pick one of five possible solutions to problems of the form A is to B as C is to ?.
Gentner [25] surveys the history of the study of analogy as a fruitful interdisciplinary convergence between psychology and AI, with significant influences from history
of science, philosophy, and linguistics. She models the use of analogy in learning
and reasoning by subprocesses: retrieval, mapping and structural alignment, evaluation, abstraction, re-representation and adaptation. Holyoak and Thagard [33] present
a theory of analogy that covers evolution of thinking in animals and in children. They
consider its implications for cognitive science in general, with examples from various
domains of human activity. Reference [56] is an interdisciplinary volume that explores
research and theory of similarity and analogical reasoning from psychological, educational, and computational perspectives. The computational perspectives deal with
problems encountered in simulating analogical processing in reasoning and problem
solving. The computational model SME in [20,24] emphasizes the role and import of
structure mapping in analogies: analogy is viewed as a mapping between structured
representations.
Indurkhya [35,36] proposes a view of metaphor as change of representation.
A key to his approach is the distinction between the concrete environment (which he
calls SMD: sensory-motor-data) and highly abstract concept networks, and the process
of interaction between these two levels that tries to preserve the autonomous structures
of both, while making the concepts correspond to the stimuli in an SMD. This key
issue, as well as other elements of his approach, are shared in some sense with the
formalism proposed in this paper. The overlaps will be pointed out at the relevant
places.
Mitchell [50] and French [22] describe two computer models of analogy-making
that model the complex interaction between perception and concepts that underlies the
creation of analogies. Hofstadter [31] chronicles efforts, including the two last models

218

Z. Arzi-Gonczarowski / Perceive this as that Analogies

as major examples, to develop computer models that provide insights into discovery,
creation, and analogical thought. A key idea that emerges from the last three cited
works is that high-level perception of situations, of patterns, and patterns of patterns
is at the roots of the cognitive processes involved.
In the machine learning literature [44] analogical inference comes under the
epithet case-based reasoning. Kolodner [38] shows the development of applications
of analogical reasoning in the AI context of case-based reasoning. A solution to
a problem can be formed by adapting a previously solved problem. One needs to
retrieve a potentially relevant case from memory and establish the correspondences
between the cases. One may then abstract the common properties of the two cases,
providing a schema that can be used for further problem solving. This is essentially
the same as Gentners subprocess of abstraction that was mentioned above. A process
in this spirit will be formalized here by a pullback categorical construction.
2.2. Human cognition and AI models
Most of the research about analogies that was cited above, including the computational models, has typically had the goal of understanding, illuminating, and imitating
human cognitive processes. For psychologists, philosophers, and linguists, the issue
of how an AI artifact produces or follows an analogy is often inseparable from the fact
that the artifact can produce or follow an analogy. Human intelligence has indeed been
a source of inspiration for research in AI, and this study is no exception. However,
this work is not committed to being empirically adequate from a human psychological
point of view.
A model is meant to represent something that can not be directly, or easily,
observed (e.g., an atom, or processes in a human mind) by another system which is
more familiar or more easily effected, and whose workings are supposed analogous to
that of the first, at least in some respects. There is selective emphasis on the features
that overlap, while ignoring other features. One might model the behavior of sound
waves upon that of waves in water, or the behavior of a gas upon that of a volume
containing moving billiard balls. However, molecules are not billiard balls, buildings
are not made of cardboard, and their miniature models are not made of concrete.
A mathematical model is typically a system of definitions, assumptions, and equations that are set up to discuss real world phenomena that are met by a physicist, an
economist, an engineer, and maybe also a cognitive scientist. Geometry emerged as a
model of the physical space in ancient Greece. From the beginning of the 17th century,
analysis developed to model motion and equilibrium of physical bodies. Mathematical probability models problems of gambling and human chance, as well as statistical
theories of mechanics and thermodynamics. There are, of course, more examples. To
apply the mathematics it is typically necessary to develop an abstract mathematical system, that approximately corresponds to the real world phenomenon. This may involve
making assumptions and simplifications. If an AI artifact is programmed according
to a formal schema (e.g., the one that is proposed here) to perceive the world and

Z. Arzi-Gonczarowski / Perceive this as that Analogies

219

analogize, its sensory-motor-neural apparatus, that interacts with its environment, will
not consist of living cells. Its workings will be similar to human cognitive workings
only in some respects, and drastically different in others. It will be argued, however,
that there is enough parallelism to call this a model of artificial perceptive cognition.
A model is a form of analogy. There are hence two distinct levels of analogy
present in this study. There is the general phenomenon of analogy making that is
being studied, and a specific analogy that is being drawn between human cognitive
processes and constructs of the proposed formal schema. Some researchers resist
to the legitimacy of the latter. Sloman [54] argues, against those whom he calls
doubters and fearers, in favor of the idea that AI artifacts may have cognition (and
even emotions). The debates about the possibility to model human cognitive processes
by formal artificial systems are, in a certain sense, analogous to some old theological
debates about attributing similar terms to man and to God. Philosophers of religion and
theologians have recognized the importance of analogy for the exposition of doctrine.
The medieval theologian Thomas Aquinas claimed that an attribute that is applied
both to man and to God is not used in a univocal sense (in the same sense), nor is
it used in an equivocal sense (totally different sense), but it is used in an analogical
sense (a similar sense). Though differences are great between man and God, there is
enough similarity that one is justified in using the same terms to apply to both. When
the same term is used, however, it must be understood that it is used in an analogical
sense when it is applied to God. Hence, yet another analogy that is suggested here
is drawn between, on one hand, the analogy that was made by Thomas Aquinas and,
on the other hand, the analogy between AI systems and the cognitive systems of their
human creators1 . When terms such as sensory-motor-neural apparatus are applied to
machines, they are used in an analogical sense. Analogy is ubiquitous, which is why
it ought to be studied.
We propose a mathematical categorical formalization of artificial perceptions as a
unified standard for the modeling of perception-related cognitive processes, analogies
and metaphors being salient examples. Ideas and concepts from research about human
cognitive processes provide pre-theoretical intuitions. They are being systematized and
modeled by mathematical context free premises. Constructs and results are inferred
from the formal premises applying methods and results from category theory, Boolean
algebra and Lukasiewiczs (three valued, non-monotonic) logic [45,46]. The suitability
of these tools to model human cognition could, indeed, be debated. They are not
suggested as the human processes themselves, but rather as parallels that can be applied
in a formal or artificial context. Whenever results are reached, it is possible to continue
the dialogue with theories of human cognition, and examine the results against these
theories. This will be done as we proceed with the constructions. It will be argued that
the resemblances noted bear relevantly on analogy-making, and that the parallelism
extends beyond the definitions. Starting from minimal definitions, more constructs are
generated, along with propositions and theorems that pertain to these constructs. If
1

Figuring out the details of this scandalous humanistic analogy is left to the reader.

220

Z. Arzi-Gonczarowski / Perceive this as that Analogies

one manages to obtain constructs and properties that would not have been obvious had
they been asserted at the outset, and they resemble certain aspects of cognition, then
this should provide additional support for the suitability of the proposed schema.
Some foundational intuitions that the proposed theory shares with studies of
human analogy-making and tries to capture and systematize, are:
High level perception (in parallel) of sensory-motor-neural impressions of cohesive
wholes in the environment provides basis for cognitive processes such as analogymaking.
Good analogies are based on structured mappings of perceptual constituents, of
patterns of these constituents, and of patterns of patterns.
An important aspect of analogy is a process of abstraction of a superordinate schema
from the analogs.
More overlaps will be pointed out at relevant points, as the presentation of the formalism proceeds.
2.3. Category theory as a tool
Category theory emphasis on mappings between structures provides a natural motivation to apply categorical tools for a formalization of analogies. Kant observed [14,
p. 66] that cognition by analogy does not signify an imperfect similarity of two things,
but a perfect similarity of relations between two quite dissimilar things. Contemporary
works about analogies and metaphors, as mentioned above, share observations that a
good analogy should be based on structured mappings. The issue of similarity between structures is fundamental in mathematical category theory, started by Eilenberg
and Maclane in [19]. Some later texts about category theory are [29,47], and even
more recently [6,7,12,16,43,52,57]. (Kant was, indeed, the one who introduced the
philosophical idea of categorization in his 18th century statement Grounding for the
Metaphysics of Morals [37].)
The term structured mapping appears frequently in studies about analogy. Some
researchers actually use variants of the terms morphisms, isomorphisms, homomorphisms: Black [10], Indurkhya [35,36], Holland and Quinn [32], and possibly others.
In most cases, however, their usage is rather loose. Indurkhyas cognitive models are
perhaps closest to the mathematical formalism that is proposed in this study. The advantages of the full category theoretical setting that is proposed here is the possibility to
rigorously integrate quite a few perceptual cognitive processes into one unified theory,
using, for example, composition of morphisms and commutative diagrams. Category
theory also provides other, more sophisticated, tools such as (but not restricted to)
natural transformations. These tools have been used in other papers [25] that build
on the same formalism. They will be discussed later.
In the general context of more formalizations of other cognitive processes,
in addition to analogies, mathematical category theory seems to be a natural candidate as well. Lawvere argues [42] that Even within mathematical experience,

Z. Arzi-Gonczarowski / Perceive this as that Analogies

221

only. . . [category]. . . theory has approximated a particular model of the general, sufficient as a foundation for a general account of all particulars. Following this theme,
category theory seems suitable for purposes of AI. This science tries, in a sense, to approximate intelligence by creating particular models of artificial intelligence as well as
foundations for a general account of such intelligence. In particular, the mathematical
categorization provides a setting for the approximation of artificial cognitive processes
by describing particular perceptual cognitive processes in particular intelligent artificial
settings, as well as foundations for a general account of intelligent artificial perceptive cognition. Marrying mathematical category theory with cognitive studies is also
proposed by Magnan and Reyes [48], who suggest that universals of the mind may
be expressed by means of universal properties in the theory of categories, and that
categorical universals constitute blueprints of the mind.
Further motivation for the application of mathematical categorical methods for
purposes of formalizing artificial perceptions and related cognitive processes has already been argued in [4,5] that introduce and discuss the mathematical premises upon
which this study is based. They also provide a summary of some representative AI
research that employs category theoretical tools. Barwise and Seligmans [8] research
can be added to that list. They have also found that the basic perspective of category
theory could be quite helpful as a guide in developing their theory of information flow
within a connected distribution system.
2.4. The category and the theory of artificial perceptions
Following the motivations to apply category theory to artificial perceptual cognitive processes, a category of artificial perceptions was introduced in [5]. Perceptions
were defined as a domain of mathematical discourse, where different perceptions represented different objects of the category. Morphisms described structure preserving
paths between perceptions. This provided tools of scrutiny for dealing with all perceptions. Structural similarities among perceptions could be studied, yet leaving ample
room for differences and variety. The categorical setting served as premises for a
mathematical theory. Its technical basis will be presented in section 3 below.
Among the applications proposed in [5] were categorical products, coproducts,
pullbacks and pushouts that were applied to formalize joint representations with varying degrees of partnership and trust. The theory was further applied for more cognitive
processes in [4], where free functors and natural transformations were applied to formalize reasonable ways to go about producing a meaningful cognitive representation
of the environment from every perception. In the present study the theory will be incremented by showing how analogical transitions can also be captured by categorical
constructs. For this purpose the basic definitions will be extended:
In the cited works the environment was fixed. The definitions allowed for a variety
of environments, however the generic representative (sub)category that was studied
consisted of perceptions of a fixed environment, with morphisms as paths between
various subjective representations and ontologies of that fixed environment. In

222

Z. Arzi-Gonczarowski / Perceive this as that Analogies

this paper this family of (sub)categories, each with a fixed environment, will be
integrated into one category. Morphisms in the unified category will also enable
transitions between perceptions that may consist of transferences and manipulations
of environments as well as transferences and changes of representations.
Structure and structure preservations in the cited works applied to the representation.
In this paper structural elements are introduced into the environment as well, and
the concept of structure preservation applies also to transitions between perceptions
of different environments.
It is noted, however, that the main import of the present paper is not in these technical
extensions to former works, but rather in highlighting common themes: The same
mathematical setting can be repeatedly employed to formalize a variety of cognitive
processes. This means that:
The premises do capture basic issues that are relevant to many cognitive processes,
suggesting a theoretical standard.
Whenever a construct or a result is repeatedly applied to more than one cognitive
capability or process, then this seems to highlight a central cognitive issue. In that
case specific references (including page numbers) to any previous mathematical
treatment, are provided.
The analysis of a variety of perceptual cognitive processes in terms of a relatively
small number of primitives predicts the possibility of tidily structured implementations with a reduced component set. Context free modules may perhaps be reusable
for different cognitive processes.
Application of mathematical methods for purposes of AI is not new [9,18]. Mathematical rigor is essential in an artificial context, because architectural requirements
need to be meticulously formalized in order to be implemented. The advantages
of mathematical formalizations as analyzed, for example, in the introduction to [17]
include clarity, precision, versatility, generalizability, testability, allowance to model
complex phenomena that are far too complex to be grasped by a verbal description,
and allowance to use results of a well developed science.
3.

The categorical premises

The essentials of the mathematical premises are presented first. (They extend
those of [4,5] by allowing transitions between perceptions of different environments.)
The abstract idea of a perception is postulated as a mathematical construct which
relates between phenomena outside the artificial agent, a set of world elements and
reflections which are internal to the artificial agent, a set of connotations.
Definition 1. A perception is a three-tuple P = hE, I, %i where E and I are finite,
disjoint sets, and % is a 3-valued predicate % : E I {t, f , u}.

Z. Arzi-Gonczarowski / Perceive this as that Analogies

223

The set E represents the environment which the machine perceives. Anything
which exists independent of the perception itself, and could perhaps be discerned by it,
could be a legitimate element of E, and hence a world element (w-element for short).
Possible example w-elements are a face, a light blow of wind, the shadow of a smile,
a slight shivering of voice, a tinge of smell or taste or color, etc. Not every sensorymotor-neural mechanism is able to discern every such outside phenomenon, and even if
it does, it may not be able to represent it internally, or attach universally conventional
connotations to them. Different perceptions might break the same reality into different
parts that are regarded as cohesive wholes. Although the external environment has an
objective existence, its organization into w-elements is subjective. It is impossible to
separate perception from the environment to which it relates. The essential role that
interaction with an environment plays in intelligence has been recognized at least since
Turing, and recently solicited by others, like Clark [15], Wells [58], and Allen [1], to
name just a few. The idea of a cognitive supraindividual that includes its environment
was also proposed by Hutchins [34]. It is also a pre-theoretical intuition of this
study that intelligence develops relative to a given, embodied, sensory-motor-neural
apparatus on one hand, combined with a given environment on the other hand. They
provide the bottom-up aspect of perceptions: from stimuli to representations.
The set I stands for the internal representation of w-elements. Its elements have
a subjective existence dependent on the machine. Anything which may be stored and
manipulated in the machine (words, symbols, icons, etc.) could be a legitimate element
of I, and hence a connotation. This provides the top-down aspect of perceptions:
from the representations to the chunking and the structuring that they impose on environments. Indurkhya [35,36] argues convincingly for the importance of the interaction
between an autonomous Reality2 which presents itself in the form of raw material
sensory stimuli, and its internal chunking and conceptualization.
The three-valued predicate % is the perception predicate (p-predicate for short)
which relates w-elements and connotations. The terminology for the various % values
is the following: If %(w, ) = t then w has connotation , if %(w, ) = f then w lacks
connotation , and if %(w, ) = u then w may either have or lack this connotation.
This undefined value might eventually become defined, but right now it is not.
Every perception has its own set of w-elements, its own set of connotations,
and its own predicative correspondence between the sets. They are given once the
instance perception is fixed, in very much the same way that the details of a cup are
accessible once perception relates to a specific instance of a cup from the category
of cups. Actual sets E and I, and the values of the p-predicate, once given, provide
a definition of a particular perception. This captures the intuition that perceptions
and sensations are innate to agents: their gestalt perceptions, mental imagery, neuralsensory-motor apparatus, function, internal organization etc. Likewise, the issue of
why the p-predicate has any one of the three values at a certain point simply warrants
no discussion: impressions are not necessarily explained.
2

Capitalized in the reference.

224

Z. Arzi-Gonczarowski / Perceive this as that Analogies

The mathematical objects stand for embodied perceptions or perceptual states.

They are high-level in the sense that they reside higher than pixels on the screen or
waveforms of sound. They are object centered and therefore happen at and above
the level of recognition of cohesive wholes, where meaning and conscious cognizance
begin to play a role. (Connectionism has been suggested as a candidate for the mechanism that underlies this level [28].)
Perceptions vary across agents, situations, etc. The flexible dynamic flow between
different high-level perceptions is formalized by perception morphisms between them:
Definition 2. Let P 1 = hE 1 , I 1 , %1 i and P 2 = hE 1 , I 2 , %2 i be two perceptions.
h : P 1 P 2 is a perception morphism (p-morphism for short) if h defines the following set mappings: h : E 1 E 2 , h : I 1 I 2 , and no-blur is the structure preservation condition: for all w E 1 , and for all I 1 , whenever %1 (w, ) 6= u then
%2 (h(w), h()) = %1 (w, ).
This definition is an extension of the definition in [4,5], which defined arrows
between perceptions with the same outside environment: E 1 = E 2 , and h : E 1 E 2
was the identity mapping.
Remark 3. The definition does not imply that both perceptions, P 1 and P 2 , exist prior
to the transition, and that the p-morphism comes later. There are cases where the
transition is creative in the sense that either one of the two perceptions gives rise to
the p-morphism, actually creating the other perception. (A path between perceptions
can be traversed backwards as well, reversing the direction of the arrow. This can
be technically effected in the mirror category, as explained in [5, p. 278].) Example
creative transitions will be mentioned later. This issue is relevant to analogizing
because, as Indurkhya [35,36] argues, some analogies actually create similarities rather
than just find pre-existing ones.
Rigid p-morphisms preserve the structure of perception in a rigid fashion:
Definition 4. h is a rigid p-morphism if, for all world elements w E 1 , and for all
connotations I 1 , %1 (w, ) = %2 (h(w), h()).
Composition of p-morphisms and the identity p-morphism are defined at the level
of set mappings, and it follows that:
Theorem 5. Perceptions with p-morphisms make a mathematical category, designated
Prc.
The proof is essentially the same as in [5, pp. 274275]. This provides a welldeveloped mathematical infrastructure for a theory of artificial perceptions, in the
same manner as the category of groups is the basis for group theory.

Z. Arzi-Gonczarowski / Perceive this as that Analogies

225

In [5, pp. 276277] and in [4, pp. 199201] example p-morphisms have already
been applied to formalize cognitive processes such as communication and interpretation
between different perceptions of the same environment. The mathematical properties
of the p-morphism (isomorphism, one to one, many to one, onto, impossible etc.) carry
meticulous information about how close these perceptions are, and pinpoints differences between them. This categorical framework has been shown to neatly capture
other cognitive processes as well. In particular, introduction of classification and mental organization cognitive processes into this sense perception framework is formalized
by free endofunctors into the subcategory of Boolean perceptions, where sets of connotations are closed under Boolean operations, namely the Is are Boolean algebras.3
These Boolean perceptions are studied in [4, pp. 207212]. Various cognitive transitions into such perceptions, using free endofunctors, are formalized and studied as well.
It is, indeed, natural for cognitive, intelligent, artifacts to sense and perceive Boolean
combinations of connotations, and to apply the associated taxonomies. If w is, for instance, a signpost, then a higher-level Boolean perception P = hE, I, %i could perceive
that %(w, small) = f , %(w, glittering ringing) = t, %(w, tasty smelly sticky) = u.
The theory provides [4, pp. 210211] categorical basis for a deductive apparatus for the
computation of the three-valued p-predicate for Boolean combinations of connotations,
and it will be employed later for analogizing.
A substantial part of [4] is dedicated to the construction of the p-morphisms that
capture high-level representation formation: Starting from basic sensory-motor-neural
perceptions and simple representations, organize and shape structured representations
of labeled impressions that can be further used for higher-level cognitive processes
(i.e., problem solving, decision making, planning, etc.). These p-morphisms map basic
represented impressions into their Boolean closures, producing logical, cognitive, representations.4 Two canonical free constructions from the category of perceptions (or
some subcategory) into the Boolean subcategory formalize reasonable ways to go about
producing a meaningful cognitive image of the environment from every perception.
One is general [4, pp. 212215], the other [4, pp. 223230] is more perceptually
acute, and it answers to a criterion of mathematical completeness and features a
categorical fixed point. Connections between analogies and these processes will be
analyzed later.
The theory thus provides embodied cognitive structures that are grounded in
authentic sensory-motor-neural impressions and basic representations of actual environments. On the other hand, they are interpretable as logical formulas, the dominant
view in AI being that the knowledge content of high level artificial reasoning processes
ought to be represented by data structures with this property [23].
In the present study the theory is incremented by showing how analogical transitions can also be captured by categorical constructs. Natural primitive constituents
3

Barwise and Seligman also introduce Boolean operations and classification into their theory of information flow [8].
4
These are creative p-morphisms in the sense of remark 3: the codomain perception does not exist prior
to the transition, but rather the p-morphism transition defines the codomain perception.

226

Z. Arzi-Gonczarowski / Perceive this as that Analogies

for perceptive analogizing between environments are provided by connotations that are
discerned by a creative intelligence. Following an acute analysis of the environment,
an observant perception may then proceed to identify a similar recomposition of (mappings of) these constituent elements in other environments, providing for an analogy.
In the following sections the process is formally analyzed within the mathematical
premises provided above.
In later sections the newly formalized processes are fused into the existing collection of cognitive processes within these premises. The unified theory provides
extensions of additional pre-theoretical conceptions, so that the mathematical formalism seems to be useful for AI: A single context free theory integrates the representation
of high-level perceptions and a variety of cognitive manipulations.
4.

An example analogy

Science teachers have traditionally used a camera as an analog to describe the

eye [26,30]. Let us first define a perception of a camera, P Cam , and a perception of an
eye, P Eye , and then define a p-morphism between the two.5 Perceptions can be conveniently described by a perception matrix, where lines represent w-elements, columns
represent connotations, and entries consist of the p-predicate value for the corresponding coordinates. In our simple example (table 1) the sets of connotations of P Cam and
P Eye are the same, so that columns are shared. The p-morphism h : P Cam P Eye
is based on the identity mapping of connotations, and h : E Cam E Eye is described by letting w and h(w) share lines: h(outside actions) = brain stem reflexes,
h(shutter) = iris, h(aperture ) = pupil, h(film) = retina, h(camera lens) = cornea,
h(lens-film distance) = eye lens + muscle. It is easy to see that h is no-blur as required by definition 2. Since h is the identity on connotations, the analogy is literal:
the similarity between w and h(w) is straight forward, connotation for connotation. The
mapping is one-to-one but not rigid (definition 4): light sensitivity and refraction have
no meaning (i.e., undefined) when applied to lens-film distance, but they are defined
(f and t, respectively) when applied to eye lens (see table 1, bottom line, columns:
light sensitive and refraction).
Remark 6. Some p-predicate values in the example could be debated. For example, if
refract is undefined for lens-film distance, one might claim that it should be undefined
for aperture as well. We shall return to such variations later, in section 8, in the context
of pullbacks.
The choice of w-elements, connotations, and the p-predicate is crucial for the
perceptions and the analogy of this example. In other contexts either the whole eyeball would make a single w-element, or, on the other hand, the sense organs and their
5

Following remark 3, this is not necessarily the order in which things occur. A very creative intelligence
might have conceived of a camera by analogy to an eye. In that case the p-morphism would have been
conceived to create its domain.

Z. Arzi-Gonczarowski / Perceive this as that Analogies

227

Table 1
Camera perception and eye perception with analogy.
I Cam = I Eye :
E Cam

Regulate
light

Light
sensitive

Adjust
focus

Refract

Dynamic
change

Control of
other object

E Eye

outside
actions

brain stem
reflexes

t/t

u/u

t/t

u/u

t/t

t/t

shutter

iris

t/t

f /f

f /f

f /f

t/t

t/t

aperture
film
camera lens
lens-film
distance

pupil

t/t

f /f

f /f

f /f

t/t

f /f

retina

f /f

t/t

f /f

f /f

f /f

f /f

cornea

f /f

f /f

f /f

t/t

f /f

f /f

eye lens
+muscle

f /f

u/f

t/t

u/t

t/t

f /f

accessories could be chunked in a different manner. Other perceptions could connote

the same w-elements with chemical properties, attributes of form and design, etc. In
these cases it would probably be harder or impossible to define an analogy. (In science,
functional resemblances are often more likely to be fundamental than qualitative ones.)
Mitchell [50] and French [22] describe computer systems that sort out the connotations and generate mappings that capture possible analogies in given environments,
thus modeling the interaction between perceptions and connotations that underlies the
creation of analogies.
Example consequences that could be derived from the analogy between the camera and the eye could be:
It may explicate the eye environment in terms of the camera environment. A perceptive intelligence that has a model of the functioning of a camera can efficiently
deduce the functioning of the eye. For example, it could deduce that the image that
is created on the retina should be inverted, because it is inverted on the film. This
could give rise to a question Do humans see everything inverted ? which is not
an unintelligent question to ask.
When a perceptive intelligence that analogizes between the camera and the eye
interacts with a human expert, this intelligence could perhaps be expected to follow
a statement like A film is never retina-sensitive.
The analogy may be used to suggest principles of functioning for similar optical
instruments. This issue will be elaborated in section 8 in the context of pullbacks.
5.

Interpretive and literal-analogical p-morphisms

In this study we are particularly interested in p-morphisms that capture cognitive

transitions between distinct environments that are similar in some analogical sense,
like the example of section 4 concerning an analogy between the camera and the eye.

228

Z. Arzi-Gonczarowski / Perceive this as that Analogies

The basic categorical tools for that transition are p-morphisms. P-morphisms can
be classified, among others, by subcategories of Prc.
Given an environment E, all possible perceptions P = hE, I, %i with that fixed E
form a subcategory PrcE . (There is actually a family of such subcategories, one for
every possible E.) PrcE is the category that was studied in [4,5]. P-morphisms in PrcE
are such that h is the identity on E (see definition 2), these are interpretive cognitive
transitions between perceptions. In interpretive p-morphisms domain connotations are
transferred to codomain connotations that represent another perception of the same
environmental phenomena.
There is a subcategory which is dual to PrcE : the generic subcategory, PrcI ,
that is introduced now. It is constructed by fixing some I, a set of connotations, and
considering all possible perceptions P = hE, I, %i with that I. P-morphisms here are
those where (in terms of definition 2) h is the identity on I. They capture transitions between different environments while fixing the internal apparatus. We shall
call these literal-analogical p-morphisms, l-analogies for short. In l-analogies domain
w-elements are transferred to codomain w-elements with the same connotations. The
analogy between the camera and the eye is formalized by an l-analogy. In [3] creative
design processes are also formalized in PrcI , where w-elements in the domain perception provide examples, similes, and raw material for the designed w-elements in
the codomain perception.
There are other possible cognitive transitions that are captured by p-morphisms
in PrcI that are not analogies in a narrow sense, because they are transitions between
w-elements that are, essentially, the same. (As Magnan and Reyes have observed [48],
categorical constructs are able to provide means to avoid over-determinations. In our
case, a p-morphism can be regarded as a generalized notion of an analogy.) Consider
a book store environment, where w-elements are books, magazines, and newspapers.
A browsing agent, with perception P = hE, I, %i, needs to know whether a given
w-element, the magazine w, is, say, the December issue. In terms of definition 1,
%(w, December issue) = u, and this perception has to be improved to the point where
that value is definite (t or f). There are three typical methods to go about this:
A pro-active store agent may perhaps take the initiative to put big red stickers that
say December Issue. In that case the environment undergoes a transition. It is
an l-analogy, h : P hE 1 , I, %1 i, that captures a manipulation of the environment
that replaces the magazine w by h(w) that has the sticker. This transition is cognitively conceived to improve browsers perceptions. It does not change Ps set of
connotations, however, it is more likely that %1 (h(w), December issue) = t.
Depending on its communicative capabilities, the agent can perhaps request the
assistance of a store agent. Possessing a different set of connotations, and, hence, a
different perception of the same store environment, the store agent gives a definite
answer that is based on its own perception of this environment, P 0 = hE, I 0 , %0 i.
As an example, P 0 may have %0 (w, from top shelf ) = t, in that case one gets an
interpretive p-morphism h : P P 0 where h(December issue) = from top shelf .

Z. Arzi-Gonczarowski / Perceive this as that Analogies

229

By the no-blur structure preservation condition on p-morphisms, whenever w is

the December issue for P, w is from the top shelf for P 0 , and whenever w is not
the December issue for P, w is not from the top shelf for P 0 . In this case the
environment is not changed, but there is an interpretation based on a transition to
another set of connotations.
The agent may peek at the dates on the cover and determine a definite (t or f) value
for %(w, December issue). In this case both the environment and the connotations
are not changed, only the perception predicate undergoes a transition, which may
perhaps be enabled by perceptual sensory-motor-neural capabilities.
The last item is captured by a pure improvement as in definition 7 below: a p-morphism
that is both in PrcE and in PrcI . There is neither an interpretation, nor an analogy. It
captures an unblurring of perception within the same environment and representation.
Definition 7. A p-morphism h : P 1 P 2 is a pure improvement if h is the identity
both on E and on I, but %1 6= %2 .
By the no-blur condition on p-morphisms, perception is more defined, namely
improved, in P 2 .
The three example transitions (i.e., p-morphisms) above have P as domain. They
are three different methods that could perhaps improve the browsing agents perception. All are formalized within the same basic simple formalism, and can be initiated
by perceptive intelligent artifacts, depending on their perceptual sensory-motor-neural
capabilities and their cognitive and communicative skills.
Similar to interpretations in PrcE , the mathematical properties of l-analogies in
PrcI (such as being an isomorphism, one-to-one, many-to-one, onto, impossible etc.),
carry meticulous information about how close the environments are, and pinpoints
differences between them. For example, transitions into larger environments could be
captured by l-analogies that are not onto. In other cases, different perceptions could
break the same environment into different w-elements, and hence cognitive transitions to larger environmental chunks could sometimes be captured by many-to-one
l-analogies. Where one perception perceives three distinct books, another perception
might perceive one trilogy. Likewise, many volumes could constitute a single encyclopedia, or a single (arbitrarily large) topic-section of books. Some interesting examples
of manipulations of environments and representations to improve perception are given
by Hutchins [34].
6.

Metaphors as factorizations of analogies

By (the extended) definition 2 of p-morphisms, a general p-morphism is both

interpretive and analogical: domain w-elements are mapped to analogical codomain
w-elements, with codomain connotations that are interpretations of the domain connotations.

230

Z. Arzi-Gonczarowski / Perceive this as that Analogies

Example 8. Consider an analogy between a simplistic perception of light and a simplistic perception of sound. For the light perception the environment consists of welements that are perceived by the light that they emit, while for the sound perception
w-elements are perceived by their sound. The light perception connotes its w-elements
by either one of {infrared, visible, ultraviolet}, while the sound perception connotes its
w-elements by either one of {infrasonic, audible, ultrasonic}. The analogy h maps:
h(infrared) = infrasonic, h(visible) = audible, and h(ultraviolet) = ultrasonic. In
this analogy neither the mapping of w-elements, nor the mapping of connotations is
the identity. It is not a literal analogy, since there is also an interpretive transfer of
connotations involved.
Example 9. Analogies that involve both transference of w-elements and transference
of connotations could be defined within the same environment and the same connotation
set, if h : E E and h : I I are not identities. Linguistic paraphrases, and also the
analogies made by the computational models of Mitchell [50] and French [22], seem
to call for such a formalization with p-endomorphisms.
6.1. Metaphors as factorizations of analogies: Technicalities
Technically, a factorization of every p-morphism into an interpretation and an
l-analogy provides a basic categorical view of the transition.
Definition 10. Let P 1 = hE 1 , I 1 , %1 i and P 2 = hE 1 , I 2 , %2 i be two perceptions, and
let h : P 1 P 2 be a p-morphism between them. An (E-I) (or (I-E)) factorization of
h consists of two p-morphisms hE , hI such that:
(1) hE PrcE and hI PrcI .
(2) Either h = hE hI (in that case it is an (E-I) factorization),
or h = hI hE (in that case it is an (I-E) factorization).
(3) The (E-I) factorization has the property that if h = g0 g00 is another factorization
that satisfies the first two requirements, then there exists a p-morphism PrcE
with hE = g0 and g00 = hI . See figure 1.
Dually, the (I-E) factorization has the property that if h = f 0 f 00 is another
factorization that satisfies the first two requirements, then there exists a p-morphism
PrcI with hI = f 0 and f 00 = hE . See figure 1.
Theorem 11. Let h be a p-morphism as in definition 10.
1. An (I-E) factorization of h is defined by (see upper part of figure 1):
I -metaphors
, (w, ) 7 (w, h()),
hI -first : P 1 P 1
I -metaphors
P 2 , (w, h()) 7 (h(w), h()),
hE -last : P 1

Z. Arzi-Gonczarowski / Perceive this as that Analogies

231

Figure 1. Two factorizations of a p-morphism.

I -metaphors

I -metaphors
= hE 1 , h(I 1 ), %1
i,
Definition of %1I -metaphors : for all w E 1 , and for all I 1 :
(
t if I such that h() = h() and %1 (w, ) = t,


%I1 -metaphors w, h() = f if I such that h() = h() and %1 (w, ) = f ,
u otherwise.

P1

2. Dually, an (E-I) factorization of h is defined by (see lower part of figure 1):

E -metaphors
, (w, ) 7 (h(w), ),
hE -first : P 1 P 1
E -metaphors
P 2 , (h(w), ) 7 (h(w), h()),
hI -last : P 1
E -metaphors
= hh(E ), I , %E -metaphors i,
P
1

Definition of
E -metaphors
%1

1
1
E -metaphors
%1
:

h(w), =

t
f
u

for all w E 1 , and for all I 1 :

if x E such that h(x) = h(w) and %1 (x, ) = t,
if x E such that h(x) = h(w) and %1 (x, ) = f ,
otherwise.

Proof. First note that, by definition, the factors are legitimate p-morphisms, because
h is. We show that the properties required by definition 10 do hold for the (E-I)
factorization of h, as illustrated in the lower part of figure 1:
1. By definition, hE -first PrcI and hI -last PrcE .
2. By definition, h = hE -first hI -last .

232

Z. Arzi-Gonczarowski / Perceive this as that Analogies

3. Let h = g0 g00 be another (E-I) factorization of h. Define PrcE by h() 7

g0 (). The mapping is legitimate since: h() = h() = g0 g00 () = g0 g00 ().
Since g00 PrcI , then also g0 () = g0 (). is no-blur by the minimality of the
E -metaphors
definition of %1
, and by g0 being a p-morphism. It is also easy to see that
0
hE -first = g and g00 = hI -last .
Proof of the required properties for the (I-E) factorization of h, as illustrated in the
upper part of figure 1, is dual.

Corollary 12. Each one of the (E-I) and (I-E) factorizations of theorem 11 is unique
up to isomorphism.
Proof. If there are two such factorizations then (or ) of item 3 in definition 10 is
a reversible p-morphism, and hence a p-isomorphism.

The intuitive idea is to gain access to the intermediate, metaphorical perceptions:
I -metaphors
I -metaphors
= hE 1 , h(I 1 ), %1
i consists of P 1 w-elements with metaphorical
P1
connotations that are inspired by P 2 .
E -metaphors
E -metaphors
= hh(E 1 ), I 1 , %1
i consists of metaphorical w-elements that are
P1
inspired by P 2 , with P 1 connotations.
Example 13. An (I-E) factorization of h of the camera-eye analogy from section 4
yields a perception of a camera with metaphorical w-elements that are borrowed from
the eye environment. One relates to a camera environment but perceives a pupil instead
of an aperture, a cornea instead of a camera lens, etc. When asked whether the eye
lens of the camera refracts, the answer is undefined. This follows from the minimal
definition of the metaphorical p-predicates in theorem 11, and from the p-predicate of
P Cam (see table 1, bottom line, Refract column, left side).
Example 14. Metaphorical terms such as white noise are based on extensions of
perceptions and analogies that are similar in spirit to example 8, where an analogy
was proposed, roughly, between visible light and audible sound, and the metaphorical
perception of a white noise can be captured by a factorization as in theorem 11.
Example 15. When linguistic translations are conceived as analogies, as proposed
in section 1, then borrowing words or phrases from another language is based on a
metaphorical transition as above.
6.2. Metaphors as factorizations of analogies: Methodological fallout
From the category theoretical point of view, the factorization is straight forward,
and the formulation of definition 10 is standard (item (3) is typically category theoretical). We wanted to state that metaphorical perceptions are based on a blend of

Z. Arzi-Gonczarowski / Perceive this as that Analogies

233

constituents from the two sides of an analogy, yet that the interfusion should be circumscribed by the perceptions involved. The categorical tools allowed a precise and
testable extension of these intuitions, that can be modeled by an artificial system more
readily than the verbal description. The metaphorical p-predicates of theorem 11
are minimally unblurred with respect to the domain perception. Technically, this
warrants compliance with item (3) of definition 10 (and also the uniqueness of the factorizations). Intuitively, item (3) of definition 10 is a systematization that captures the
circumscription of metaphors: they cannot become arbitrarily far-fetched. A specific
example of the way that this limitation works was shown above for the camera-eye
analogy: The metaphorical perception of the eye lens of the camera leaves undefined
the issue of whether it refracts. The metaphor is not carried beyond certain bounds,
and these bounds are rigorously defined by the formalism. This is one more instance
where the category of artificial perceptions provides us with a reward for the technical
efforts.
The corresponding metaphorical transitions are formalized by the respective
p-morphisms, hI -first and hE -first . Categorical properties of these p-morphisms can
be applied to evaluate the metaphorical transition: Is it one-to-one? Is it rigid (definition 4)? A positive answer to both last questions means, for example, simpler
metaphors. If a p-morphism h is already in PrcE (or in PrcI , as in the camera-eye
analogy of section 4), then the relevant factor is the identity. This could perhaps
be allusive of Nietzsches observation that literal truth is merely dead or fossilized
metaphor.
The proposed factorization is relevant to other perceptualcognitive issues. One
of them is the evasive boundary between literal meaning, on one hand, and methaphorical or metonymic meaning, on the other hand. Philosophers as well as linguists vary
in their views. Lakoff, for example, holds the position that many of our conventional everyday concepts are metaphorical. Indurkhya, on the other hand, holds the
position that to qualify as a metaphor, a concept needs to be given an unconventional interpretation. The proposed factorization may be applied to bridge the gap
between the views: An agent may start out with a perception such as P 1 of theorem 11. Applying an analogical transition h into another perception, like P 2 of
theorem 11, and then factorizing that analogy, the agent may perceive with the implied metaphors. If the agent internalizes the intermediate metaphorical perception,
E -metaphors
E -metaphors
= hh(E 1 ), I 1 , %1
i from figure 1, and gets used to employing
say P 1
it on a regular basis, then the metaphors become conventional or even literal. By
internalization it is meant that in subsequent recognition or usage of the metaphor the
agent does not go through the entire process of transition and factorization all over
again, but rather that the metaphorical perception becomes a basic perception, with the
same status that P 1 had had at the outset. Metaphors may start out as unconventional
and become conventional. The degree of conventionality is in the eye of the perceiver, depending on how much computational effort (i.e., mappings, p-morphisms,
factorizations) is invested in finally identifying one of the three values {t, f , u} with
instances of %E1 -metaphors (w, ). Perceptual states, either individual or shared by a soci-

234

Z. Arzi-Gonczarowski / Perceive this as that Analogies

ety of agents, undergo continuous dynamic transitions, so that both conventionality

and unconventionality are subjective and fluid.
Another perceptualcognitive problem that could perhaps benefit from the proposed factorization of p-morphisms is the long recognized fact that many concepts
(or words) can take on an infinite number of different senses, because there is no
limit on the number of contexts. Clark [15] argued that context contributes to the
meaning of a concept, on each particular occasion of use, in a deep way. Others
have made related points. The sense variant of the concept is constructed at the time
of perception from some core meaning, in combination with the context in which it
occurs. All agree that, although meanings change from agent to agent, and they also
change within agents over time and experiences, some critical invariable aspect of
meaning is held by individual perceptions and must be shared in order for perceptual
transitions to make sense. One issue that remains open for rigorous formalization
is how to fix a core invariant meaning across environments, individual representations, and instances of perceptual transitions. The structure preservation condition
on p-morphisms binds different representations through environments, and analogical environments through representations. The factorization may untangle the knot.
Loosely, if one fixes the interpretation (i.e., hI is the identity), then one can perform a meaningful analogical transition between environments. Dually, when fixing
the environment (i.e., hE is the identity), one can perform a meaningful interpretive transition between representations. Complex transitions can be broken (i.e.,
factorized) into (recomposable) smaller steps of this kind, thus, perhaps, grasping
the evasive invariable aspect of meaning with a loose and flexible, yet durable, harness.

7.

Structural alignment: Analogizing with structured arrows

At the roots of powerful analogies and metaphors lie more than just pointwise
mappings. Research background about the significance of structured analogies was
mentioned in section 2 from several perspectives. Thagard [55, p. 81] argues that
. . .Two analogs are similar to each other at a superficial level if they involve similar
concepts. . . However, powerful analogies involve not just superficial similarities, but
also deeper structural relations.
The proposed categorical formalism will be refined now to capture structural
features of analogies. Perceptible structure consists essentially of observed relations
among the constituents of a perception. They play an important role in the conception
of analogies, and imply restrictions on the manner in which things can be mapped
by an analogy. Based on a given perception P = hE, I, %i, the formalism provides
tools for capturing observation of lawlike patterns among w-elements of E as well as
among connotations of I. We start by introducing a quasi ordering on w-elements.
Definitions 16 and 17 below are novel in the proposed categorical setting.

Z. Arzi-Gonczarowski / Perceive this as that Analogies

235

Definition 16. Let P = hE, I, %i be a perception. A w-element x E is subjacent to

another w-element y E, designated x E y, if, for all connotations in I, %(x, ) = t
implies that %(y, ) = t and %(y, ) = f implies that %(x, ) = f .
Definition 17. Let P = hE, I, %i be a perception. Two w-elements x, y in E are congeneric if, for all connotations in I, %(x, ) = %(y, ).
It is easy to see that congenerics are subjacent one to the other, and that the
congeneric relation is an equivalence relation. The subjacency relation is a quasi
ordering on E, and a partial ordering on the equivalence classes of the congeneric
relation. Example subjacencies of w-elements in P Cam are, for instance, aperture E
shutter, and also shutter E outside actions. Example subjacencies of w-elements in
P Eye are, for instance, pupil E iris, and also iris E brain stem reflexes. These patterns
are, indeed, meaningful for understanding the structure and the functioning of the
camera and the eye.
Parallel relations between connotations were already introduced and discussed in
[4, pp. 203207]. They are repeated below for easy reference.
Definition 18. Let P = hE, I, %i be a perception. A connotation I subsumes
another connotation I, designated E , if, for all w-elements w in E, %(w, ) = t
implies that %(w, ) = t and %(w, ) = f implies that %(w, ) = f .
Definition 19. Let P = hE, I, %i be a perception. Two connotations , in I are
synonyms if, for all w-elements w in E, %(w, ) = %(w, ).
It is easy to see that synonyms subsume one another, and that synonymity is an
equivalence relation. The subsumption relation is a quasi ordering on I, and a partial
ordering on the equivalence classes of synonyms. Perceptions that are unique up to
synonyms were studied in [5, pp. 291295]. In a dual manner environments that are
unique up to congenerics can be defined, mutatis mutandis. The intuitive idea is that
more duplicates of the same constituent do not feature a significant difference in the
structural aspect of a perception.
In the perceptions P Cam and P Eye from table 1, one can easily observe several
subsumptions of connotations. Subsumptions that hold in both perceptions are, for
example, control of another w-element E dynamic change, light amount regulation E
dynamic change, focus adjustment E dynamic change. These patterns are, indeed,
meaningful for understanding the functioning of the camera and the eye.
Hence, introduction of a structural element both into the internal representation I,
on one hand, and into the environment E, on the other hand, occurs when a perceptive
intelligence observes lawlike patterns between the constituents of its perception. (The
definitions are inspired by the definitions that Lukasiewicz gave to the biconditional
and to the conditional in his 3-valued logic [45,46]. The choice is discussed in [4,
pp. 192194].)

236

Z. Arzi-Gonczarowski / Perceive this as that Analogies

The subjacency and subsumption relationships do not necessarily represent causal

relationships, or isa relationships, although these are particular possibilities. Observation of a lawlike pattern is independent of an explanation pertaining to why or
how the pattern holds6 . The observation may, indeed, stimulate reasoning higher-level
processes to track the causes or the nature of the pattern, but this is a separate cognitive
process. Understanding, for example, how or why the iris is subjacent to brain stem
reflexes is a complicated issue that may be studied separately.
Intelligence is typically marked by a discerning perception and understanding
of its environment, and analogy making ought to benefit from these capabilities and
demonstrate them. When a perceptive intelligence observes lawlike patterns in its
environment, it would be natural to incorporate the observations into its cognitive
processes. The correlate in a category theoretical setting is to preserve the structure
when applying a categorical construct. The definition of monotonicity from [4, p. 206]
is extended now to include also the structure that was defined on the environment E
(definitions 16 and 17).
Definition 20. Let P 1 = hE 1 , I 1 , %1 i and P 2 = hE 1 , I 2 , %2 i be two perceptions.
A p-morphism h : P 1 P 2 is monotone if, for all , in I 1 , E implies that
h() E h(), and, for all x, y in E 1 , x E y implies that h(x) E h(y).
In the monotone subcategory of perceptions arrows are restricted to monotone
ones, and hence all categorical constructs preserve the involved structure. Rigid pmorphisms as in definition 4 are always monotone. Non-monotonicity occurs only (but
not necessarily) when there is some unblurring of perception. When x E y, and for
some I it so happens that %(x, ) = %(y, ) = u, then if h(x) has the connotation
h() but h(y) does not, then this is a typical case of non-monotonicity, since h(x)
is not subjacent to h(y). In the case where an analogy is described by a monotone
p-morphism, this analogy is systematized by a mapping that preserves structure, and
so are the metaphors that are implied by the factorizations of a monotone analogy.
In the l-analogy between the camera and the eye, some subsumptions of connotations and some subjacencies of w-elements are preserved. Among them are the
examples that were provided above following the definitions. The example subsumptions of connotations are literally preserved: control of another w-element E dynamic
change, light amount regulation E dynamic change, focus adjustment E dynamic
change. The example subjacencies of w-elements are also preserved by the analogy:
aperture E shutter and also pupil E iris, shutter E outside actions and also iris E brain
stem reflexes. This structural alignment actually means that the implied structure and
functioning of the camera and the eye are similar, making the analogy between the
camera and the eye a meaningful analogy.
The bad news is that the p-morphism that describes the l-analogy between the
camera and the eye is, however, not monotone. As mentioned above, non-monotonicity
6

Perception follows, in this context, the scientific enquiry principle hypotheses non fingo [11, Newton,
p. 261].

Z. Arzi-Gonczarowski / Perceive this as that Analogies

237

may occur only when there is some unblurring of perception. Hence the row of
table 1 that should be examined carefully is the bottom row, namely subjacencies
that involve either the lens-film distance or the eye lens+muscle, and subsumptions
that involve either light sensitivity or refraction. There is, indeed, non-monotonicity
there: lens-film distance E outside actions but eye lens 6E brain stem reflexes because
%(eye lens, refraction) = t, but %(brain stem reflexes, refraction) = f . It follows that
one has to omit the refraction connotation or accept an analogy that is not monotone.
The principal difference between a camera and an eye has to do with the process of
focus adjustment, and the proposed formal setting shows sensitivity to this difference,
providing additional support that it is an effective formalism. The mapping of lens-film
distance to the flexible eye lens is not rigid (definition 4). In particular, the l-analogy
is not a p-isomorphism.
What happened? Non-rigidity, and the resulting non-monotonicity, occurred exactly at the point where the analogy is not perfect. When an analogy is captured by
a p-morphism, then the monotonicity of that p-morphism happens to be a touchstone
for the uniformity of the analogy, and the pigeon-hole of non-monotonicity provides
meticulous information about the point where the analogy slips7 . P-morphisms,
monotonicity and rigidity are precise, applicable and testable tools of analysis that are
enabled by the categorical formalism.

8.

Analogizing with pullbacks

The analogy between the camera and the eye still feels to be a good one. Let
us see if the categorical tools are capable of systematizing the following argument: In
both systems there is focus adjustment, which may be performed in different ways.
What one actually claims in this argument is that if one avoids over determination,
by not specifying exactly how focus is adjusted, then at that level of perception, the
camera and the eye are analogous.
The cognitive process that underlies this argument consists of first generalizing
the two perceptions into a third, abstracted, perception of a generalized image generating system. Such a system consists of abstracted w-elements that should be able to
produce an image on a light sensitive medium while performing refraction, adjusting
focus, regulating light, etc. The camera and the eye become particular substitution
instances of this general schema. The generalized perception has an abstract environment that corresponds to what Glynn et al. [26] entitle a superordinate system,
Thagard [55] entitles analogical schema, and Gentner entitles the generation of that
generalized perception as the subprocess of abstraction. The loose verbal description
can be captured by well developed tools of category theory. This will be done now.
7

The term slippage was coined by Hofstadter and his colleagues [22,31,50].

238

Z. Arzi-Gonczarowski / Perceive this as that Analogies

Figure 2. Start by matching everything with everything: p-product.

8.1. Schema abstraction with p-products and p-pullbacks

The definitions of p-product and p-pullback was introduced in [5, pp. 288296]
for the subcategory PrcE . The definitions are now extended to the entire category
Prc. A product of two perceptions is their minimal change common blur. Loosely,
each perception is blurred exactly to the point where there is no conflict with the other
perception.
Theorem 21. The product of P 1 and P 2 is the perception
P 1 P 2 = hE 1 E 2 , I 1 I 2 , %1 %2 i,
where the set of w-elements is the set product of the sets of w-elements, the set of
connotations is the set product of the sets of connotations, and the p-predicate %1 %2
is defined as follows:
(
t if and only if i = 1, 2 %i (wi , i ) = t,


%1 %2 (w1 , w2 ), (1 , 2 ) = f if and only if i = 1, 2 %i (wi , i ) = f ,
u otherwise.
The associated projections are (see figure 2) i : P 1 P 2 P i , where i (1 , 2 ) = i
and i (w1 , w2 ) = wi .
The proof that this is indeed a categorical product, and that the p-product always
exists, is essentially the same as the proof in [5, pp. 289290]. In the subcategory
PrcE all p-morphisms identity-map E to itself, so that the product environment E E
may be degenerated into E.
In the general case, most of the p-predicate values in a p-product are going to
be undefined, since most pairs should consist of two essentially different coordinates.
For example, if one constructs the p-product of the perceptions of the camera and
the eye: P Cam P Eye , then a juxtaposition of, say, outside actions and retina, is
meaningless. There is no similarity between them. This total absence of similarity
is captured formally by the undefined p-predicate value for all pairs of connotations
(1 , 2 ): %1 %2 ((outside actions, retina), (1 , 2 )) = u. The undefined value can be
easily computed from the values in table 1 and from the definition of %1 %2 in
theorem 21.

Z. Arzi-Gonczarowski / Perceive this as that Analogies

239

Figure 3. Selecting analogs.

The intuitive idea behind p-products is to:

1. First try matching (i.e., analogizing) everything with everything by juxtaposition
of all possible pairs in the cartesian products E 1 E 2 and I 1 I 2 .
2. The next step will be to sort out only the pairs that feature perceptible connaturalness, unlike the pair (outside actions,retina), and more like the pairs that share
rows in table 1.
The p-product construction systematizes the first step above. A p-pullback will systematize the second step (see figure 3).
Before going into the rigorous formalism, an informal description might help:
Pointing out a subset of w-elements and a subset of connotations in a perception (e.g.,
in the p-product) can be simply regarded as an inclusion p-morphism: one-to-one and
rigid. This happens to be the definition of a p-equalizer as defined for PrcE in [5,
p. 285]. An extension to the entire category Prc is straight forward: A p-equalizer
is a one-to-one and rigid p-morphism. The selection provides a new perception that
offers possible analogs because it consists of selected pairs from the p-products that
feature similarity. (It still remains to define what one means exactly by pairs that
feature similarity.) The selection process is systematized by the p-equalizer h in the
diagram of figure 3.
Definition 22. Let P 1 and P 2 be two perceptions. A superordinate perception that
they generate8 , P 1 uP 2 P 1 P 2 , consists of w-elements (wk1 , wk2 ) and connotations
(j1 , j2 ) from the product, that feature perceptible connaturalness as follows:
(1) For all connotation pairs (j1 , j2 ), and for all w-element pairs (wk1 , wk2 ), either
i = 1, 2, %i (wk , j ) 6= f , or i = 1, 2, %i (wk , j ) 6= t.
(2) Let P 1 u P 2 = hE, I, %i.
For every connotation in this perception there exists at least one w-element
w in this perception, such that %(w , ) 6= u.
Likewise, for every w-element w in this perception, there exists at least one
connotation w in this perception, such that %(w, w ) 6= u.
8

There are probably cases of perceptions that can generate more than one superordinate perception.

240

Z. Arzi-Gonczarowski / Perceive this as that Analogies

Item (1) warrants that the coordinates of the selected pairs are not dissimilar.
Technically, this condition is necessary and sufficient for the p-pullback theorem 24
below. For example, the pair that was mentioned before, (outside actions, retina),
cannot be included in the superordinate perception P Cam u P Eye , because for all the
connotations {regulate light, adjust focus, dynamic change, control of another
object}, %Cam (outside actions, ) = t but %Eye (retina, ) = f , in contradiction with the
condition of item (1).
Item (2) requires the existence of a more definite, perceptible, similarity between
coordinates. A superordinate perception P Cam u P Eye , that can be generated by the
perceptions of the camera and the eye, may consist of w-elements that are the pairs
that share lines in table 1, namely: (outside actions, brain stem reflexes) (shutter, iris),
(aperture, pupil), (film, retina), (amera lens, cornea), (lens-film distance, eye lens +
muscle). It is easy to verify that they answer both conditions.
By definition of the p-product p-predicate, the condition of item (2) ensures that
for every connotation pair (j1 , j2 ) there exists at least one w-element pair (wj1 , wj2 )
such that i = 1, 2, %i (wji , ji ) = t or i = 1, 2, %i (wji , ji ) = f . Intuitively,
in the first part of item (2), the superordinate w-element w = (wj1 , wj2 ) suggests a
constituent that underlies the connaturalness of j1 and j2 . This constituent may
justify compromises that might be introduced by weaker similarities (or, rather, nondissimilarities) of other constituents that comply only with the first condition. The
second part of item (2) is the dual requirement for analogical w-elements. For example,
in a superordinate perception P Cam u P Eye , that might be generated by the perceptions
of the camera and the eye, the connotation adjust focus is a constituent that underlies
a certain connaturalness of lens-film distance and eye lens + muscle. This justifies any
compromises that are introduced by weaker similarities (or, rather, non-dissimilarities).
They have to do with whether either one of lens-film distance or eye lens + muscle
has the connotations light sensitivity and refraction.
Remark 23. Whether or not it is enough to analogize with w-elements and connotations
that comply only with negative definite p-predicate values (f ) in item (2) is hard to
answer at a global categorical level. This is related to issues that are discussed by
Goodman in [27]. It is possible to restrict the definition to the existence of w such
that %(w, w ) = t, and to w such that %(w , ) = t, because, from the technical point
of view, item (1) of the definition is already necessary and sufficient for the pullback
construction of theorem 24 below.
Now, from the category theoretical point of view, a combination of a product
and an equalizer yields a pullback, as shown in figure 4. The advantages of the
formalization by a pullback follow from the categorical property of pullbacks, which
is illustrated in figure 5, and defined as follows:
Given arrows f1 : P 1 Q and f2 : P 2 Q, the commutative diagram of figure 5,
where g1 , g2 have the same domain, and g1 f1 = g2 f2, is a pullback, if it has

Z. Arzi-Gonczarowski / Perceive this as that Analogies

241

Figure 4. Looking for analogs: p-pullback.

Figure 5. Categorical pullback.

the property that any other commutative diagram (where g10 f1 = g20 f2 ) can be
completed by a unique .
One import of the pullback property to the process of schema abstraction that is being
systematized here, is that the pullback perception is minimally blurred: Connotations
of w-elements in the superordinate perception that need not be blurred remain defined.
The abstract schema generalizes, but avoids determination only there where this is
absolutely necessary (e.g., in the case of whether lens-film distance or eye lens+muscle
have the connotations light sensitivity and refraction).9
Theorem 24. Let P 1 and P 2 be two perceptions. A superordinate perception as in
definition 22 can be effected by a categorical pullback construction.
Proof. For every pair that should be introduced into the superordinate perception,
define the p-morphisms fi of the diagram in figure 4 to map the coordinates of the
pair to the same element of (a trivially chosen) Q. The first condition of definition 22
is necessary and sufficient for that, so that the fi s should comply with the no-blur
condition of definition 2. The superordinate perception is then defined by the ppullback of f1 , f2 , that is categorically defined by the p-equalizer of (1 f1 , 2 f2 ).
It consists of the desired subsets of pairs from the product P 1 P 2 .

9

Readers interested in category theoretical workings are invited to compare the formulation of the
pullback property with the formulation of the minimality of metaphors in item (3) of definition 10.

242

Z. Arzi-Gonczarowski / Perceive this as that Analogies

Figure 6. Schema abstraction for the Camera and the Eye.

A superordinate perception that can be generated by the perceptions of the camera

and the eye with connotation pairs as described above consists of w-elements that are
the pairs that share lines in table 1. The connotations of the superordinate perception
are pairs of identical connotations, so that pairs can be degenerated into the original
single connotation. (Actually, this is a construction in the subcategory PrcI where I
is fixed, so the construction is based on set products of environments only.)
It is already known from the former section that weak similarities (or, rather,
non-dissimilarities) are to be expected from the bottom row of table 1. Indeed,
the only cases of weaker similarity are associated with the w-element pair w =
(lens-film distance, eye lens + muscle). The following p-predicate values answer only
to the weaker condition in item (1) of definition 22: %i (wi , light sensitive) 6= t,
and %i (wi , refraction) 6= f . However, the following p-predicate values answer also
to the stronger condition in item (2) of that definition: %i (wi , adjust focus) = t,
%i (wi , dynamic change) = t, %i (wi , regulate light) = f , and %i (wi , control of other
object) = f , featuring a definite kind of similarity. There is, hence, a schema abstraction that captures the idea of a w-element that changes dynamically and takes care of
focusing (and does not have to do with light regulation and control of other objects).
There are projection l-analogies from this superordinate perception to the camera
perception: g1 = h 1 : P Cam u P Eye P Cam , and to the eye perception: g2 =
h 2 : P Cam u P Eye P Eye , as shown in figure 6. The p-predicate for (lens-film
distance, eye lens+muscle) with refraction and with light sensitivity unblurs in different
ways for each of the projecting l-analogies, providing a formal account of the difference
between the camera and the eye.
8.2. Properties of the superordinate perception
In section 7 the concept of structural alignment in analogies was systematized
within the categorical framework by the monotonicity of the relevant p-morphisms.
A natural question that may be asked now is what happens to the structure when a
superordinate perception is generated by a p-pullback as above. In particular, we ask
whether the projection l-analogies from the superordinate perception to the concrete
perceptions (that gave rise to the whole process) are monotone. The answer is positive,
as shown in the following:

Z. Arzi-Gonczarowski / Perceive this as that Analogies

243

Theorem 25. Let P 1 u P 2 be a superordinate perception.

If (, 0 ) and (, 0 ) are connotations in P 1 uP 2 such that E in P 1 and 0 E 0
in P 2 , then also (, 0 ) E (, 0 ) in P 1 u P 2 .
If (x, x0 ) and (y, y 0 ) are w-elements in P 1 u P 2 such that x E y in P 1 and x0 E y 0
in P 2 , then also (x, x0 ) E (y, y 0 ) in P 1 u P 2 .
Proof. The theorem is proven using the definition of p-products from theorem 21. For
example, assume negatively that for some in P 1 u P 2 , (x, x0 ) has the connotation
but it is undefined whether (y, y 0 ) has the connotation (in which case (x, x0 ) E (y, y 0 )
does not hold). In that case, by definition of the product p-predicate, either x E y
does not hold in P 1 , or x0 E y 0 does not hold in P 2 , in contradiction to the negative
assumption.

The import of the last theorem is that the superordinate perception captures not
only the similarities between constituents, but also the common structure. Moreover,
even in the presence of a weaker similarity between constituents, lawlike patterns, if
they are shared, will not be erased.
Corollary 26. The projection l-analogies from the superordinate perception (the
p-pullback) g1 = h 1 and g2 = h 2 are monotone.
In the case of the superordinate perception P Cam u P Eye , for example, the schema
abstraction got rid of the problematic structure that could not be aligned by the analogy.
The mathematical formalism tells us in a precise and testable manner that the generalization was exactly enough for that, neither more, nor less. This is yet another instance
where one reaps the fruits of the effort invested in the formalism of the category of
artificial perceptions.
Of course, the pullback construction could yield an empty superordinate perception, or a perception with very few constituents. Theorem 25 tells us that the process
would erase significant structure if it pertained exclusively either to P 1 or to P 2 .
This would essentially mean that the analogy highlights a similarity with respect to
few aspects only. In that case one should either de-emphasize the other aspects, or
simply accept the fact that these perceptions are not similar enough for a meaningful
analogy to be drawn.
An ideal analogy is almost too good because it means, in a sense, that the two
sides of the analogy are essentially the same. An analogy between two really different
things typically slips somewhere. P-pullbacks in the proposed category of artificial
perceptions are formal tools that can be practically applied for the definition and
analysis of analogies that are good but have slips. They formalize a cognitive process
of abstraction that extracts the similarity (both in single features and in structure) and
separates it from incidental differences.
Another advantage of analogizing with p-pullbacks vs. analogizing with simple
arrows is the advantage of symmetry. A p-morphism means that there is a domain

244

Z. Arzi-Gonczarowski / Perceive this as that Analogies

that maps into a codomain, and the roles are not symmetric. With a p-pullback there
is no restriction about one perception being consistently less blurred or more general
than the other. For example, in remark 6 of section 4, it was mentioned that some
p-predicate values of the example in that section could be changed. The symmetrical
construction of a pullback is more robust to variations of this nature. They should not
cause an essentially meaningful analogy to collapse.
There is an additional intuition about analogies and cognitive processes that is
formalized by superordinate perceptions and p-pullbacks. In section 5 it was mentioned
that the analogy between the camera and the eye may be used to suggest principles of
functioning for other optical instruments. Analogizing with superordinate perceptions
and p-pullbacks provides basis for a cognitive process where analogies are used to
suggest hypotheses about a more general rule or setting. The superordinate perception
of an image generating system suggests principles of functional design for other
optical instruments: It suggests that such systems are designed by a composition
of constituents that refract, adjust focus, etc. This could serve as basis for either
analogical designs on one hand, or, on the other hand, for reasoning about optical
instruments. When faced with an existing optical instrument, or with a prospective
design of such an instrument, a perceptive intelligence that has gone through the
perceptual cognitive process that is described by the p-pullback, would perhaps be able
to analogize further and look for a refracting modul, for the control hierarchy, etc. Such
cognitive processes are typically based on analogies with past perceptual experience,
reinforced by generalizing cognitive mechanisms. A variety of analogs is accumulated
and unified into a single superordinate analog, providing the additional advantage of
avoiding the pitfall of fixation on a single previous analog. The superordinate analog
internalizes the idea that focus adjustment, for example, can be achieved in various
ways.
Indeed, the more general the setting, the less detailed will be the superordinate
perception. If the construction has already pulled back many perceptions of various
optical instruments, it may end up with a general observation that the only essential constituent of an optical instrument either refracts or reflects. A combination of
generality and specificity could be attained by restriction of the perceptions that are
incorporated into the superordinate perception, such as perceptions of optical instruments with focus adjustment.
The projection l-analogies from the superordinate perception to the concrete perceptions that generated the p-pullback can be factorized as in section 6. In that case it
would perhaps be more appropriate to entitle the constituents of the intermediate perceptions as specific examples rather than metaphors. There is considerable overlap
between terms such as examples, similes, models, metaphors, analogs etc. The purpose
of this paper is to formally capture the essence of the underlying cognitive processes
and to highlight structural similarities between them.
P-products and p-pullbacks are more complex categorical constructs than basic
p-morphisms. This seems to capture the intuition that a generalization of two concrete
systems into an abstract system requires more cognitive effort than the construction

Z. Arzi-Gonczarowski / Perceive this as that Analogies

245

of a simple analogy by one, straight forward, p-morphism. This parallelism between

cognitive effort and the complexity of the formal construct was not asserted at the
outset. Therefore it provides additional support for the suitability of the proposed
schema.
To summarize, analogizing with superordinate perceptions and p-pullbacks has
the following advantages:
It formalizes detection of an essence of similarity, and of structure, if it exists, while
offering the compromise that blurs and de-emphasizes differences.
It formalizes a cognitive process of generalization and abstraction, suggesting the
existence of general common principles that underlie the perceptions that are offered
as analogs.
It features a formal symmetry between the two analogs.
The formalism opens the way for programmable procedures that might be capable
of detecting similarities between perceptions and deriving analogies. The generation of
the p-product, and also the comparison and selection process that yields the p-pullback
are quite straight forward. Two relevant issues are worth mentioning:
A formalism is not an algorithm. There is indeed room for more research on
complexity and implementational issues. The categorical construct just offers a
basic foundational theoretical standard.
Following the discussion in section 2.2, the proposed formalism is not meant to be
empirically adequate from a human psychological point of view. It is not claimed
that human analogy-making actually works by generating a p-product and then a
p-pullback, it probably does not. It is often the case in AI that artificial systems
typically apply tedious search (as in scanning the entire p-product) where humans
would typically conjure specialized effective retrieval methods. Every intelligence
does whatever it is good at doing. Computer vs. human chess playing, a landmark
success of AI, is probably such an example. It is likely that the internal workings of
Deeper Blue were not identical to those of Kasparov. However, they did interact in
the same game and it is generally accepted that their capabilities may be legitimately
compared. Fusing the discussion in section 2.2 with the ideas of this section,
the analogy that was suggested between human cognitive processes and artificial
systems is roughly sketched in figure 7, where the arrows are not isomorphisms.

9.

Most structured arrows for analogies

More complex structures in perceptions are lawlike patterns (i.e., subjacencies

and congenerics, and also subsumptions and synonyms) that can be observed between
Boolean combinations of w-elements (and also of connotations). A Boolean combination of connotations can be understood intuitively, as explained in section 3.

246

Z. Arzi-Gonczarowski / Perceive this as that Analogies

Figure 7. The categorical model.

Example 27. In the perception of the eye, the iris regulates the amount of light
is NOT light sensitive:


% iris, light amount regulation (light sensitivity) = t.

AND

An adequate deductive apparatus for the computation of the three-valued

p-predicate for such Boolean combinations is provided in [4, pp. 210211].
A Boolean combination of w-elements is less intuitive, because it often involves
an abstractive conception of environmental elements that do not really exist. However, it can be technically defined in the same way. The same deductive apparatus that
was just mentioned for the computation of the three-valued p-predicate for Boolean
combinations of connotations can be applied for the computation of the three-valued
p-predicate for Boolean combinations of w-elements, mutatis mutandis. Loosely, disjunction of two w-elements is conceived as another w-element that should have the
union of the connotations that they have, and should lack the intersection of the connotations that they lack. Conjunction of w-elements is dual: it is conceived as another
w-element that should have the intersection of the connotations that they have, and
should lack the union of the connotations that they lack. A negated w-element is a
w-element that has the connotations that the original w-element lacks, and lacks the
connotations that the original w-element has. Negated w-elements do seem less intuitive because the inversion of truth values alienates the simile. However, even in
the human context negative examples are acceptable in many domains, and are often
considered better than no example at all (e.g., See this? this is exactly what you
do not want to do. . .).
For example, by (retina) one designates an imaginary w-element. It has connotations that are the negations of the connotations of retina, namely it positively features
light amount regulation, focus adjustment, refraction, dynamic change, and control of
other object, and it lacks light sensitivity. Actually, it is congeneric to a disjunction of
iris and eye lens:
I Eye

%Eye (retina, ) = %Eye (iris eyelens , ).

This kind of Boolean structure, that consists of lawlike patterns (in the example above:
congenerics) also between Boolean combinations of constituents, is defined and studied
in [4, pp. 225227], where p-morphisms that preserve this structure are also defined.

Z. Arzi-Gonczarowski / Perceive this as that Analogies

247

This is done for the subcategory PrcE , and can be easily extended in a straight forward manner to the entire category Prc. Introduction of the Boolean structure into
the internal representation I is systematized by a suitable free endofunctor into the
subcategory of Boolean perceptions [4, pp. 228230]. Introduction of this structure
into the perceived environment E can be dually systematized by a suitable free endofunctor, mutatis mutandis. The import for analogy making, and cognition in general,
is that the same perceptual acuity, which is captured by Boolean structure, pervades all
cognitive processes: interpretations, communications, mental representations, analogy
and metaphor making, as well as other cognitive processes that will be mentioned in
section 11 that describes subsequent research.
If the Boolean structure is preserved by an analogical p-morphism, or by a superordinate perception, then the analogy is even more structured, and so are the implied
metaphors. Theorem 25 obviously holds also for this extended notion of structure.
Example 28. In the perceptions of the camera and the eye it can be observed that, for
all w-elements w, %(w, light amount regulation focus adjustment) = %(w, dynamic
change), meaning that dynamic change is synonymous to (light amount regulation
focus adjustment). This is a lawlike pattern of Boolean combination of connotations
that is preserved by the l-analogy.
Some lawlike patterns of Boolean combinations of w-elements in the analogy between the camera and the eye are preserved as well. It can be observed
that (film camera lens) E (outside actions). Namely, an imaginary w-element
with all connotations except for light sensitivity and refraction should have been
subjacent to outside actions. This subjacency is preserved by the analogy, since
(retina cornea) E (brain stem reflexes): an imaginary w-element with all connotations except for light sensitivity and refraction should have been subjacent to brain
stem reflexes. These patterns are meaningful to the understanding of the functioning
of the camera and the eye, and their preservation is, hence, meaningful as well.
In section 3 it was explained how introduction of classification and organization
cognitive processes into the proposed sense perception framework is formalized in the
subcategory of Boolean perceptions that have sets of connotations that are closed under
Boolean operations, namely the Is are Boolean algebras. A substantial part of [4]
was dedicated to the construction of the p-morphisms that map basic neural-sensorymotor perceptions and simple representations into their Boolean closures, producing
logical, cognitive, representations. Out of the two canonical free constructions into
the Boolean subcategory that formalize generation of meaningful cognitive images of
the environment, the one that is more perceptually acute, and answers to a criterion
of mathematical completeness, is based on the internalization of lawlike patterns (also
between Boolean combinations) as described above. This provides a connecting thread
between various perceptual-cognitive processes of discerning perceptions. (The association is at all possible because all these processes share the same mathematical, context
free, premises from section 3.) Both the generation of a perceptually acute cognitive

248

Z. Arzi-Gonczarowski / Perceive this as that Analogies

image of the environment, as well as the generation of highly structured analogies,

are based on internalizations of lawlike patterns between (Boolean combinations of)
perceptual constituents. There is hence a connection between the capability to provide
and fathom insightful analogies and the capability to construct a subtle cognitive image of the environment. The key to both processes is the internalization of as many
perceptually observable patterns as possible.
A conclusion with both formal and applicational import is that a single component that should be capable of a certain mental acuity the detection of lawlike
patterns between (Boolean combinations of) perceptual constituents is a significant
and reusable component for various, seemingly different, intelligent perceptual cognitive processes. A programmed implementation of such a component is described
in [13], where they describe a logic-based methodology for analyzing observations
that is based on the detection of hidden patterns in the data. Combinations of such
patterns are used for developing general classification procedures. We propose that
similar methodologies can be applied by an intelligent artifact to embodied ontologies,
and that the classifications be used for more cognitive activities. With a neat distinction
among subjective perceptions, external environments, and internal representations, the
mathematical categorical toolkit of the theory of artificial perceptions provides basis
for employing the Boolean methodology for a broader spectrum of cognitive activities.
10. Cognitive science revisited
In section 2 it was suggested that one should return to theories of human cognition
in order to examine the results of this study. This dialogue between the artificial-formal
and the natural-human contexts is resumed to tie the ends of our intuitions. Indeed,
ideas such as:

Metaphors are implied by analogies . . .

One may generalize from analogs . . .
Subtle, acute perceptions provide insightful analogies and metaphors . . .
Subtle, acute perceptions classify and organize . . .

. . . are quite intuitive and not new. The mathematical infrastructure that is provided
by the definition and study of the category and the theory of artificial perceptions has
enabled us to extend, to systematize, and to formalize these intuitive pre-theoretical
conceptions in a parallel, categorical, setting. By employing well-developed tools of
mathematics, it is thus possible to model complex perceptual cognitive phenomena that
could be grasped only loosely by verbal descriptions. The mathematical categorical
framework models them in a precise, testable and applicable form, augmenting the
evidence from [4,5] that mathematical categorization of artificial perceptions can be
useful to AI.
There is an additional intuition about intuitions that is extended by the parallelism. A prudent study of analogies should be conducted with a reservation. The

Z. Arzi-Gonczarowski / Perceive this as that Analogies

249

fruitfulness of analogies depends on whether any testable consequences can be deduced

from them, which is likely to depend on whether the resemblance is fundamental or
superficial. However, when analogies are employed, it is not always possible to show
that the resemblances noted bear relevantly on the points to be established, whereas
the differences are irrelevant. It is not always possible to support the analogy with
independently established considerations, and it is not always the case that if things
are alike in some ways, they will be alike in others, hence analogies should not be
carried too far.
With all these well known limitations, researchers still often view analogizing
as a fundamental paradigm of cognitive capabilities, analogy following and analogy
making being hallmarks of intelligence tests. Fauconnier [21], for example, examines
the mappings that link mental spaces as a central component of meaning construction.
He argues that the same principles operate at the highest levels of scientific, artistic,
and literary thought, and at the lower level of elementary understanding. He counts
analogical mappings as a key cognitive operation, along with conceptual integration
and blending, discourse management, induction, and recursion. Another example is
the analysis of metaphor as a special conveyer of meaning, and as a rich cognitive
device, that is carried by Ross [53]. He links the study of metaphor processing to
wider issues in cognitive science.
The proposed theory of artificial perceptions may offer its parallel clue to the paradigmatic significance of analogy-making. Arrows are, indeed, the most fundamental
component of categorical constructions. A competence in fabricating arrows is thus
indicative of a potential capability to erect all other categorical constructions as well.
In the proposed categorical formalism many cognitive processes are systematized by
p-morphisms. It was already argued in section 5 that a p-morphism can be regarded as
a generalized notion of an analogy. It follows that an intelligent perceptive artifact that
is competent at analogizing, holds a fundamental key to other categorical constructions
of the theory, namely to other perceptual cognitive processes that are formalized by this
setting. This parallelism between the significance of arrows in the categorical setting,
and the significance of analogies in human cognition, provides an additional supporting
argument that the proposed theory provides a fruitful and effective formalization of
intuitions about perceptual cognitive processes, namely that the analogs drawn here
(between the artificial-formal and the natural-human contexts) are not superficial, and
that they bear relevantly on perceptual cognitive processes.
11. Summary and subsequent work
It is often the case in AI and in cognitive science, and also in category theory
and in logic, that one applies a methodology while studying it. In this work an analogy has been drawn between perceptual cognitive processes of analogy and category
theoretical constructs. It was shown how quite a few properties and structural features
of cognitive processes are preserved and systematized in the categorical setting. Using

250

Z. Arzi-Gonczarowski / Perceive this as that Analogies

the terminology of Magnan and Reyes [48], the consequences of this analogy provide
basis for blueprints that can be followed in an artificial perceptual cognitive context.
Subsequent research, some of which builds on the constructs that were introduced
in this paper, is concerned with further cognitive processes that can be captured by
the same mathematical categorical framework. Cognitively conceived creative design
processes are formalized by natural transformations of environments in [3]. This offers
additional parallelisms concerning the connection between reasoning processes, design
processes, and analogies. Another research direction [2] is concerned with an extension
of the formalism to derive from within, in a categorical manner, emotive reactions that
may be triggered by perception (a shift in the focus of attention is an example). This
provides premises for the incorporation of behavior, reactions, conflicting reactions,
prioritizing, and eventually also affective elements, into the cognitive processes that
are captured by this framework. This should offer a formalization of the interfusion
of the sensible and the sensitive aspects of embodied artificial intelligence. Since all
the perceptual cognitive processes are based on the same premises, they can form an
integrated schema of affective artificial cognition. Integration, in the categorical sense,
is the eventual construction of a commutative diagram that interweaves the processes.
This study suggests a theoretical standard against which analogies and related
perceptual cognitive processes can be generated and analyzed in an artificial intelligent
context. It does not provide a programmed or an algorithmic implementation. However, the analysis in terms of the relatively small number of primitives of the proposed
category (w-element, connotation, perception predicate), supported by category theoretical constructions, predicts the possibility of tidily structured implementations with
a reduced component set. Context free categorical procedures, such as p-morphisms,
seem to be reusable for different cognitive processes. It was also shown how a context free mental component that detects lawlike Boolean patterns of perceptual constituents could be reusable for various artificial cognitive activities.
References
[1] J.F. Allen, AI growing up the changes and opportunities, AI Magazine 19(4) (1998) 1323.
[2] Z. Arzi-Gonczarowski, Wisely non rational a categorical view of emotional cognitive artificial
perceptions, in: Papers from the 1998 AAAI Fall Symposium: Emotional and Intelligent The
Tangled Knot of Cognition, ed. D. Canamero, Orlando, FL, October 1998 (AAAI Press, 1998)
pp. 712.
[3] Z. Arzi-Gonczarowski, Categorical tools for perceptive design: Formalizing the artificial inner eye,
in: Computational Models of Creative Design IV, eds. J.S. Gero and M.L. Maher, Key Centre of
Design Computing and Cognition, University of Sydney, Australia (1999) pp. 321354.
[4] Z. Arzi-Gonczarowski and D. Lehmann, From environments to representations a mathematical
theory of artificial perceptions, Artif. Intell. 102(2) (1998) 187247.
[5] Z. Arzi-Gonczarowski and D. Lehmann, Introducing the mathematical category of artificial perceptions, Ann. Math. Artif. Intell. 23(3,4) (1998) 267298.
[6] A. Asperti and G. Longo, Categories, Types, and Structures (MIT Press, Cambridge, MA, 1991).
[7] M. Barr and C. Wells, Category Theory for Computing Science (Prentice-Hall, Englewood Cliffs,
NJ, 1995).

Z. Arzi-Gonczarowski / Perceive this as that Analogies

251

[8] J. Barwise and J. Seligman, Information Flow, Cambridge Tracts in Theoretical Computer Science,
Vol. 44 (Cambridge University Press, 1997).
[9] E.A. Bender, Mathematical Methods in Artificial Intelligence (IEEE, Los Alamitos, CA, 1995).
[10] M. Black, Metaphor. in: Philosophical Perspectives on Metaphor, ed. M. Johnson (University of
Minnesota Press, Minneapolis, MN, 1981) pp. 6382. A reprint of an article from 1955.
[11] S. Blackburn, Dictionary of Philosophy (Oxford University Press, 1996).
[12] F. Borceux, A Handbook of Categorical Algebra (Cambridge University Press, 1993).
[13] E. Boros, P.L. Hammer, T. Ibaraki, A. Kogan, E. Mayoraz and I. Muchnik, An implementation of
logical analysis of data, Rutcor Research Report RRR 22-96, Rutgers University, New Brunswick,
NJ (July 1996).
[14] H. Caygill, A Kant Dictionary (Blackwell Publishers, Great Britain, 1995).
[15] A. Clark, Being There, Putting Brain, Body, and World Together Again (MIT Press, Cambridge,
MA, 1997).
[16] R.L. Crole, Categories for Types (Cambridge University Press, 1993).
[17] M.A. Croon and F.J.R. Van de Vijver, eds., Viability of Mathematical Models in the Social and
Behavioral Sciences (Swets and Zeitlinger B.V., Lisse, 1994).
[18] E.R. Doughherty and C.R. Giardina, Mathematical Methods for Artificial Intelligence and Autonomous Systems (Prentice-Hall, Englewood Cliffs, NJ, 1988).
[19] S. Eilenberg and S. Mac Lane, General theory of natural equivalences, Trans. Amer. Math. Soc. 58
(1945) 231294.
[20] B. Falkenhainer, K.D. Forbus and D. Gentner, The structure-mapping engine, Artif. Intell. 41(1)
(1990) 163.
[21] G. Fauconnier, Mappings in Thought and Language (Cambridge University Press, 1997).
[22] R.M. French, The Subtlety of Sameness A Theory and Computer Model of Analogy Making,
A Bradford Book, (MIT Press, Cambridge, MA, 1995).
[23] M. Genesereth and N.J. Nilsson, Logical Foundations of Artificial Intelligence (Morgan Publishers,
San Mateo, CA, 1987).
[24] D. Gentner, The mechanisms of analogical learning, in: Similarity and Analogical Reasoning, eds.
S. Vosniadou and A. Ortony (Cambridge University Press, 1989) pp. 199241.
[25] D. Gentner, Analogy in: A Companion to Cognitive Science (Blackwell, 1998) chapter II(1),
pp. 107113.
[26] S.M. Glynn, R. Duit and R.B. Thiele, Teaching science with analogies: A strategy for constructing
knowledge, in: Learning Science in the Schools: Research Reforming Practice, eds. S.M. Glynn
and R. Duit (Lawrence Erlbaum, Mahwah, NJ, 1995) chapter 11.
[27] N. Goodman, Fact, Fiction, and Forecast (Harvard University Press, 4th edition, 1983).
[28] S. Harnad, The symbol grounding problem, Phys. D 42 (1990) 335346.
[29] H. Herrlich and G.E. Strecker, Category Theory (Allyn and Bacon, 1973).
[30] P.G. Hewitt, Conceptual Physics The Highschool Physics Program (Addison-Wesley, Menlo Park,
CA, 2nd edition, 1992).
[31] D. Hofstadter, Fluid Concepts and Creative Analogies Computer Models of the Fundamental
Mechanisms of Thought (Basic Books, 1995).
[32] D. Holland and N. Quinn, eds., Cultural Models in Language and Thought (Cambridge University
Press, Cambridge, MA, 1987).
[33] K.J. Holyoak and P. Thagard, Mental Leaps Analogy in Creative Thought, A Bradford Book (MIT
Press, Cambridge, MA, 1995).
[34] E. Hutchins, Cognition in the Wild (MIT Press, Cambridge, MA, 1995).
[35] B. Indurkhya, Metaphor and Cognition (Kluwer Academic, 1992).
[36] B. Indurkhya, Metaphor as change of representation: an artificial intelligence perspective, J. Exper.
Theor. Artif. Intell. 9(1) (1997) 136.
[37] I. Kant, Grounding for the Metaphysics of Morals (Hackett Publishing Company, Inc., Indianapo-

252

[38]
[39]
[40]
[41]
[42]

[43]
[44]
[45]
[46]
[47]
[48]

[49]
[50]
[51]
[52]
[53]
[54]

[55]
[56]
[57]
[58]

Z. Arzi-Gonczarowski / Perceive this as that Analogies

lis/Cambridge, 1992). Translated by J.W. Ellington. Originally published in 1785 as Grundlegung
zur Metaphysik der Sitten.
J. Kolodner, Case-Based Reasoning (Morgan Kaufman, San Mateo, CA, 1993).
G. Lakoff, The death of dead metaphor, Metaphor and Symbolic Activity 2(2) (1987) 143147.
G. Lakoff, Women, Fire, and Dangerous Things What Categories Reveal About the Mind (The
University of Chicago Press, 1987).
G. Lakoff and M. Johnson, Metaphors We Live By (The University of Chicago Press, 1980).
F.W. Lawvere, Tools for the advancement of objective logic: Closed categories and toposes, in: The
Logical Foundations of Cognition, eds. J. Macnamara and G.E. Reyes, (Oxford University Press,
1994) pp. 4355.
F.W. Lawvere and S.H. Schanuel, Conceptual Mathematics (Cambridge University Press, 1997).
D. Leake, ed., Case Based Reasoning: Experiences, Lessons, and Future Directions (AAAI Press,
Menlo Park, CA, 1996).
J. Lukasiewicz, Many-valued systems of propositional logic (1930), in: Polish Logic, ed. S. McCall
(Oxford University Press, 1967).
J. Lukasiewicz, On 3-valued logic (1920), in: Polish Logic, ed. S. McCall (Oxford University Press,
1967).
S. MacLane, Categories for the Working Mathematician (Springer-Verlag, 1972).
F. Magnan and G.E. Reyes, Category theory as a conceptual tool in the study of cognition, in: The
Logical Foundations of Cognition, eds. J. Macnamara and G.E. Reyes (Oxford University Press,
1994) pp. 5790.
M. Minsky, ed., Semantic Information Processing (MIT Press, Cambridge, MA, 1968).
M. Mitchell, Analogy Mapping as Perception A Computer Model, A Bradford Book (MIT Press,
Cambridge, MA, 1993).
D.J. Moser, Sze-chuan pepper and Coca-Cola: The translation of Godel, Escher, Bach into Chinese,
Babel 37(2) (1991) 7595.
B.C. Pierce, Basic Category Theory for Computer Scientists (MIT Press, Cambridge, MA, 1991).
D. Ross, Metaphor, Meaning and Cognition (Peter Lang, 1993).
A. Sloman, Architectural requirements for human-like agents, both natural and artificial (what sorts
of machines can love?), in: Human Cognition and Social Agent Technology, ed. K. Dautenhahn,
Advances in Consciousness Research (John Benjamins Publishing, forthcoming).
P. Thagard, Mind (MIT Press, Cambridge, MA, 1996).
S. Vosniadou and A. Ortony, eds., Similarity and Analogical Reasoning (Cambridge University
Press, 1989).
R.F.C. Walters, Categories and Computer Science (Cambridge University Press, 1991).
A.J. Wells, Turings analysis of computation and theories of cogitive architecture, Cognitive Sci.
22(3) (1998) 269294.