Submitted to the JOURNAL OF COGNITIVE SYSTEMS RESEARCH
*********** COGSYS-RS-99-02-(book review), KANERVA, 26 JULY 99 ***********


            Constructive Approach to Understanding Minds:
            A Review of Stan Franklin's ARTIFICIAL MINDS

                            Pentti Kanerva
             RWCP Theoretical Foundation SICS Laboratory
                   Real World Computing Partnership
                Swedish Institute of Computer Science
                  Box 1263, SE-164 29 Kista, Sweden
                       e-mail: kanerva@sics.se


ARTIFICIAL MINDS, Stan Franklin, MIT Press, Cambridge, Massachusetts,
1995, xi+449 pp., ISBN 0-262-06178-3 (HB), 0-262-56109-3 (PB).


Stan Franklin's book ARTIFICIAL MINDS explores different ways of
understanding minds.  It surveys philosophies of mind, signs of mind
in the animal kingdom, the use of computers for modeling life and
mind's functions, brain activity that could underly mind, robots as
embodiments of minds, and the debate and controversy surrounding
different points of view; and it considers the possibility of
building systems with artificial minds.  The book is appropriate for
an educated reader with an interest in such matters as well as the
professional, thanks to its style, scope, and depth.  After you have
read it once and absorbed the material, read it again for the insight
that comes from seeing the work as a whole.  It is well worth it.

Our popular notion has it that a mind is something particularly human,
exemplified by conscious thinking, feeling, emotions, reasoning,
intuition, remembering, communicating.  Things relating to the mind
are called mental and are contrasted with the physical; mind being
contrasted with matter.  Against this picture the idea of artificial
minds may seem radical, even heretical.  Not only would bears and
birds and butterflies have minds but we could build machines with
minds, which is more controversial than machines with intelligence.

Perhaps artificial intelligence is more acceptable to us than
artificial minds because we have more of a handle on intelligence--we
measure it crudely with IQ tests, for example--and because the field
of Artificial Intelligence, or AI, has been around longer.  For more
than 30 years AI has been taught in leading universities giving it
legitimacy even if most of us don't know how AI works.  It is a little
like our accepting computers without knowing how they work.  How many
of us could build a computer from delay lines and logic gates, from
relays and switches?  Is it even necessary to know computers at that
level of detail, unless you are an engineer?  I believe that it is
doubly so, if we are to understand artificial minds: first because the
computer is at the core of many models of mind, as shown in Franklin's
book, and then to drive home the idea that when we are looking for an
underlying mechanism of X, we are looking for something very unlike X;
computers allow us to see and to reflect on complex and obscure
behavior arising from mechanisms that we fully understand.  Such
relating of behavior to mechanism is central to Franklin's ARTIFICIAL
MINDS.

The study of the mind has historically been the domain of philosophy
and more recently of the HUMANITIES, which Webster (10th Collegiate)
defines as "the branches of learning (as philosophy, languages)
that investigate human constructs and concerns as opposed to natural
processes (as physics or chemistry)."  Again we see the mental
contrasted with the physical: mind as not resulting from natural
processes.  However, early Western philosophy included also physical
and biological science--Aristotle felt qualified to do it all--that
is, before the mechanisms of astronomy, physics, chemistry,
metabolism, and heredity, in roughly that order, were known anything
like how we know them now.

Western thought is marked by a passage from ignorance through
philosophy into natural sciences, as more and more complex phenomena,
in roughly the above order, have been explained in terms of underlying
mechanisms.  Franklin's book follows this trend by taking us to the
next frontier: to explaining minds in terms of underlying mechanisms.
Undoubtedly there are frontiers beyond this one, such as those studied
in the social sciences and frontiers yet unknown, but this one is
special.  We have arrived at a point where the thing being explained
is just as complex as the thing doing the explaining; where the mind
is trying to sort out its own mechanism.

Is this asking for the impossible, like lifting ourselves by our
bootstraps?  Although this is fertile ground for philosophical debate,
and Franklin reviews some of it for us, the issues are not resolved
by debate and thought experiments alone.  We must also gather facts
and evidence on how minds work and fail.  But even that is tricky when
we ourselves are so completely immersed in the phenomenon we try to
understand.  Just think of the difficulty that humankind has had in
seeing ourselves as minute specks on a tiny planet rotating around its
axis and whirling around the sun in a huge and indifferent universe,
when we experience ourselves as being at the center of it all, on
stationary Earth that provides for our needs, with the heavens
revolving around us.  The illusion is irresistible--because it is our
reality.  It is even more so with minds.  Each person's mental life
is that person's reality, and what gives rise to it--what underlies
mental life--is hidden from a person's mind.

THIRD-PERSON VIEW.  How does Franklin navigate the "illusions"
arising from deep self-involvement?  Mainly by insisting that mind
is not uniquely human but a product of evolution, so that the animal
kingdom offers us countless varieties and degrees of mind.  Simple
animals have simple minds, animals most like us have minds most
like ours, and an infant's simple mind develops into an adult's more
complex mind.  This gives us a third-person view of mind and turns the
basic question about minds into: When should we regard something as
having a mind?  Rather than trying to define it sharply by necessary
and sufficient conditions, we take the human mind as a standard and
humanlike behavior as an indicator of mind: the more like us something
acts, the more of a mind it has.  This also is the logic behind the
test that Turing has proposed for deciding whether a machine can
think.

In Franklin's (third-person) view, the working of a mind is seen in
action or, more precisely, in the INTERACTION of a thing with its
environment.  The question then becomes: How is it accomplished?  How
does sensing lead to action?  What internal structures and functions
are indicated?  The underlying premise is that organized activity of
some kind within the individual is necessary.  With humans and animals
we take it to be the activity of the brain.  This approach leads into
examining physical structure--the organization of matter--as a
prerequisite of mind and, finally, into the possibility of artificial
minds.

ENGINEERING VIEW.  Franklin classifies the study of mind according
to two criteria, top-down vs. bottom-up, and analytic vs. synthetic.
He sympathizes most strongly with the bottom-up, synthetic approach,
which he calls "Mechanisms of Mind"; the book is indeed an
exploration of the mechanisms of mind.  According to it, we take
simple components whose working we understand and from them design
a system that works so like a mind that it surprises even its
designer--in other words, it has lifelike emergent behavior.
On this, Franklin quotes Carver Mead: "If we really understand
a system we will be able to build it.  Conversely, we can be sure
that we do not fully understand a system until we have synthesized
and demonstrated a working model."  I call this the ENGINEERING
VIEW and will expand on it.

Engineering does not rank highly in intellectual debate--nowhere
near philosophy or art or mathematics or the nature of consciousness.
Like agriculture, it seems utilitarian, mundane, intellectually
uninteresting.  Why bother with neurons and synapses, or with delay
lines and logic gates, when we are talking of minds?  And that
exactly is the reason!  Minds are exceedingly difficult for us to
grasp because their underpinnings are invisible to us, and so we talk
about minds in high-level, abstract terms such as beliefs, desires,
intentions, motivation, plans, and goals.  This gives us a feeling
that we really understand minds, but when we try in AI to design
minds from such abstract building blocks, we are not particularly
successful.  According to the engineering view, we are far from
understanding minds.  Thus, the engineering view is an exacting
standard against which to judge the depth of our understanding.
It disciplines our minds and fosters intellectual honesty and, as
things stand, it gives us cause to be humble about the depth of our
understanding.

In yet another way engineering can lead to understanding minds.
Brains are too complex--there are too many details--for us to
sort out, and duplicating a brain does not equal understanding it.
However, the PRINCIPLES by which brains accomplish their feat need
not be many and complicated, they just are effective and non-obvious.
Once the principles are understood, we can build systems based on
them and thereby learn how minds work.

AUTONOMOUS-SYSTEM VIEW.  Franklin views minds as control systems
of autonomous agents.  They allow agents to learn, and so we need
a plausible account of learning.  Learning is sometimes likened to
programming, with the brain corresponding to computer hardware and
the mind to software.  However, many of our programs for mindlike
functions are based on our abstract understanding of minds.  In doing
so, we build OUR meanings into the system and thereby become a part
of the system and lose a third-person view of it.  By definition,
an autonomous agent has its OWN meanings--Franklin expresses it
by saying that minds create information--and anything resembling
a program comes into existence through the system's own actions and
interactions with its environment.  This agrees with the Enactive
Paradigm of mind put forth by Varela, Thompson, and Rosch and reviewed
sympathetically by Franklin.  From raising children we know that
explicit instruction, which is the mind's equivalent of being
programmed from without, is inefficient compared to learning from
example.  It shows that natural minds learn very differently from
today's programmed computers, so that equating the (human) mind with
software is more like a cliche than a helpful analogy.  I doubt that
minds can be truly appreciated and understood without understanding
also the hardware.

The topic of MEANING brings us to looking at symbolic representation
through the eyes of biologists, engineers, and roboticists.  Human
language is a paradigm of symbolic representation: words are arbitrary
patterns for meanings, and the meanings are something else somewhere
else.  Contrast this with the following.  An autonomous system is
coupled to its environment by sensors and effectors; in an animal it
is by huge arrays of neurons, and everything about the world that is
available to the animal's mind is defined by neural firing patterns.
We can think of these patterns as the animal's symbols, except that
they are not arbitrary.  Some patterns make the animal move, some
are pleasurable, some painful, and so forth, and we can no longer
distinguish symbol from meaning.  Similarly with artificial autonomous
agents, some patterns over some of the gates and delay lines are the
agent's own meanings: there is no representation in the traditional AI
sense.  This idea is taken to its extreme in Brooks' Nouvelle AI and
Subsumption Architecture, in which Franklin detects the seeds of the
Third AI Debate.

The book describes all three.  The first AI debate was over the
possibility that a machine, a computer, could think.  The second
pitted connectionist AI against traditional symbolic AI, the
connectionists rallying over the idea that artificial neural nets
are more brainlike than computers, and the others arguing that it did
not matter.  The third AI debate is about the necessity and nature
of mental representations.  The old and the nouvelle AI would have us
in opposite corners, when the interesting action is more likely to
be all over the playing field.  Franklin hints at the possibility of
varieties and degrees of symbolism, akin to varieties and degrees of
mind.  I agree with that view so much so that it defines my present
research.  It seems to me that arbitrary symbols are not needed for
the things we do as infants to satisfy our basic needs, whereas
our making up, telling, and understanding of stories, for example,
implies internal modeling of the world and sophisticated use of more
or less arbitrary symbols.  The internal modeling could be partly
traditional symbolic AI-ish but it cannot be fully so, and so we must
look for representations that allow for both arbitrary and meaningful
symbols, and everything in between.  Pollack's RAAM and Chalmers'
experimenting with it, which Franklin reviews, are steps in that
direction.

Through the book runs a philosophical undercurrent that leads to
Franklin's Action Selection Paradigm of mind, which is summarized in
the following three statements (p. 419): (1) Cognition is the process
by which an autonomous agent selects actions; (2) actions emerge from
the interaction of multiple, diverse, relatively independent modules;
and (3) a cognitive system functions adequately when it successfully
satisfies its needs within its environment.  That is not meant as a
recipe for engineering but as criteria by which to gauge philosophies
and models of mind.  They are to be judged from the viewpoint of an
autonomous system that interacts with its environment, and the success
criterion means that things with minds could come to be by evolution.

In addition to controlling action, the human mind is occupied with all
sorts of things for their own sake: we sing and dance and chat with
friends and produce plays and go to the theater and watch the sunset
for the pleasure, stimulation, and peace of mind that they give us.
There must be more to minds than what meets the third person's eye.
My reading of ARTIFICIAL MINDS suggests that Franklin would agree.

Franklin's notion of mind does not necessarily culminate in the
human mind.  Some people--Franklin cites Moravec, and Margulis and
Sagan--speculate that once mind's mechanisms are understood we will
build silicon minds more powerful than ours.  We don't even have to
go beyond what already exists.  A family, a community, a corporation,
a nation, and all of humanity can be thought of as autonomous agents
with superminds that are harder yet for us to comprehend than an
individual human mind.  Such superminds pose an interesting puzzle:
When the idea of mind as a society of agents is taken to its logical
conclusion, it obscures rather than clarifies the notion of mind and
mind's mechanisms.

Franklin's writing is admirable.  It is thoughtful, lively,
informative, and accessible.  He explains the working of more than
a dozen models.  I already knew some of them well--one being my
own--and Franklin's descriptions of them gives me faith in the
quality of instruction I received on the others.  He shows concern
for the readers' difficulties with unfamiliar subject matter, and so
he provides background material, is generous with diagrams, figures,
and tables, avoids specialized language, and steers around ambiguity
with examples.  It is a precious gift to the reader not to have to
struggle with the presentation when struggling with new concepts and
points of view is challenging enough.  I noticed only a few instances
where his watchfulness gave way to his long and intimate familiarity
with the subject matter.  To characterize the mind (or language or
consciousness or free will) as "scalar" or not "Boolean," or to
reject the notion of a "LINEAR continuum" of minds (p. 412) is
mathematicians' shorthand that can distract others.

The book is very helpful when you want to learn more.  Numerous
comments throughout the text, together with 14 pages of references,
give a wealth of pointers to further reading.  To the reading
list I would like to add BRAINS, BEHAVIOR, AND ROBOTICS by James
Albus (Peterborough, N.H.: Byte Books, 1981), NEURONS AND SYMBOLS:
THE STUFF THAT MIND IS MADE OF by Igor Aleksander and Helen Morton
(London: Chapman & Hall, 1993), and Aleksander's IMPOSSIBLE MINDS:
MY NEURONS, MY CONSCIOUSNESS (London: Imperial College Press, 1996).
All three take an engineering view and an autonomous-system view
similar to Franklin's.

The production of a book is an endless battle against trivial errors,
and we rarely claim total victory.  There are a few wrong words that
spelling programs didn't catch; in figures 11.3-11.5 the arrows for
successor and predecessor links have been interchanged; some page
numbers in the index are off by one; and 20 bifurcations would give
Moravec's bush robot (Fig. 15.3) a mere million cilia rather than a
trillion--did he mean 20 "tetra-furcations"?  The reader can easily
disregard such errors.

I will conclude with a personal view on the state of our science as
evident in the observations, experiments, and models that Franklin
writes about.  Our models, both symbolic and connectionist, are
abstract at too high a level.  They are more like metaphors for mind
than basic mechanisms, they act more like mirrors than microscopes.
They are generalizations from what we see and introspect of mind's
working.  The problem with such abstraction is that even when it
credibly describes behavior, it does not constrain sufficiently the
underlying mechanisms.  It does not tell how brains and their models
should be constructed, and so we get very little in terms of emergent
properties from our models, very few pleasant surprises.  Our models
hardly begin to explain how brains learn and minds develop.

Since the underlying mechanisms are hidden they have to be inferred
and then tested my modeling, and so the power of our minds to imagine
and to conjecture is crucial.  That power comes from experience.
We need to know what psychology and neuroscience can tell us about
the mind's working, and we need to know mathematical systems.  For
example, a mathematical model of a certain kind will be suggested
and understood only by someone who is familiar with that kind of
mathematics.

But does our understanding of computers, or minds, have to go down
to the level of delay lines and logic gates, or neurons and synapses?
In the following sense it does.  They make it plain how unlike the
infrastructure is from the phenomenon we want to understand, being
it the behavior of a computer or a mind.  These components and their
actions are meaningless in the sense that the whole system behaves
meaningfully.  So the ORGANIZATION of the components is crucial and
is what we must understand.  Notice also that delay lines and logic
gates refer to principles of operation--namely, to holding and
combining of data, holding and combining of patterns--except that
these principles are not mysterious to us the way minds are, and they
are easily realized in physical devices.  I see a close analogy to
understanding chemistry in terms of atomic structure, or life forms
in terms of chromosomes, genes, and the genetic code, and I bet that
we will never give them up for the old ways of thinking about matter
or heredity.

Constructive modeling starts with simple components, which are then
built into circuits for working with patterns.  The patterns realize
abstract states, and the circuits govern state transitions, accounting
for the system's behavior.  Some patterns or states are meaningful
through grounding, for which Brook's subsumption architecture is a
model, and others become meaningful by composition, which in some
ways must resemble symbolic processing.  All of this, too, is abstract
but it is abstract at a low level and the high-level abstractions--new
concepts--are based on the low-level ones.  The notion that
implementation does not matter is wishful thinking for a scientist.

Major discoveries and hard work lie ahead before we uncover a
foundation for the working of the mind that is anything like
chromosomes, genes, and the genetic code are for the working of
heredity and the evolution of life.  Yet we must try, or accept
the alternative that minds work by magic.  ARTIFICIAL MINDS is
an excellent introduction to the ways of our trying.


*********** COGSYS-RS-99-02-(book review), KANERVA, 26 JULY 99 ***********