|
ARTIFICIAL INTELLIGENCE: IS IT POSSIBLE? Rob Harle (c) 1999 T The purpose of this essay is an attempt to determine, in principle, if it is possible to create a conscious machine. I will not discuss the moral nor ethical issues such an entity would raise, nor the desirability of such an enterprise. I also do not want to discuss the minute details of AI (Artificial Intelligence) methods - programming, computer architecture and so on. Apart from the fact that I’m not qualified to do so, getting bogged down in detail can often obscure or confound the overall picture. Also this essay is more of an exploration of the issues involved and a general introduction to the rather esoteric field of AI, in which it seems there are as yet no definitive authorities.
The notion of creating an artificial entity, that is, “a human
created in the image of a human” has been around a lot longer than the
last thirty or forty years of the Computer Age.
Mary Shelley’s “Frankenstein” dealt with the ‘scientific’
creation of a human-like entity on many levels.
One of these levels was socialisation and in an indirect, somewhat
ironic way, is relevant to the possible creation in the early twenty first
century of an intelligent, conscious silicone - digital - molecular
‘creature’. I discuss the socialisation aspect of AI in detail further
on.
The creation of
computational devices and those with a ‘memory’ goes back many
thousands of years. The
Chinese invented the Abacus some 5,000 years ago, the water clock around
3,000 years. The first truly
mechanical clock, which ticked, was built in 725CE.
In 1642 Pascal invented the world’s first automatic calculating
machine, the Pascaline. Shortly
after in 1694 a computer was invented by Leibnez.
Weaving was to
see many mechanical and computational innovations at the beginning and
during the Industrial Revolution. The
Jacquard loom, around 1805, was operated by punch cards, this was a
significant ‘advance’ in transferring instructions from a human to a
machine to produce a product, in absentia. This was of course the
precursor to punch card computers of the mid-nineteen hundreds.
Charles Babbage
developed the Analytic Engine, the world’s first ‘real’ computer.
Ada Lovelace published her own notes on this engine and
“speculated on the ability of computers to emulate human
intelligence”. The date was
1843! Hollerith perfected the
automatic punch card tabulating machine and founded a company which later
became IBM.
In 1921 the word
“robot” was coined by Czech dramatist, Capek. His science fiction
drama “Rossum’s Universal Robots” describes how intelligent
machines, originally servants of humanity, end up taking over and
destroying their creators.
In 1937 Turing
developed his famous Turing Machine, a theoretical model of a computer.
In 1950 he developed his almost infamous, “Turing Test” to
assess the intelligence of a machine compared with that of a human.
The usefulness of this test is now seriously doubted. (see Searle
and Hofstadter). From about
this time on computers simply became faster and with greater and greater
capacity and problem solving capabilities (Kurzweil, 1999. pp.261-280.
Relays gave way to vacuum
tubes, these gave way to transistors which in turn gave way to the
silicone chip. This
ubiquitous ‘chip’ may soon give way to a new technology known as
molecular organic circuitry. Silicone
chips are reaching maximum efficiency due to physical limitations, a
custom-designed molecule called rotaxane, may replace the old style
digital silicone computer with a molecular computer within five years
(Chang, 1999.) This very brief history helps ground the following discussion
and shows the progressive nature of human invention. In a sense it helps
justify the predictions of futurists, such as Kurzweil, that machines will
be conscious and equally as intelligent as humans within thirty years (Kurzweil,
1999).
Before I address
the question of artificial consciousness we need to have a clear idea of
what constitutes ‘natural’ human consciousness. This is no easy task, which is evidenced by the monumental
amount of literature devoted to the “race for consciousness”.
Searle defines consciousness as, “...a biological feature of
human and certain animal brains. It
is caused by neurobiological processes and is as much a part of the
natural biological order as any other biological features such as
photosynthesis, digestion, or mitosis” (Searle, 1992. p.90).
Hobson believes,
“...the mind is all the information in the brain” and defines
consciousness as, “...the brain’s awareness of some of that
information” (Hobson, 1994. pp.202-204).
Hobson also believes that the brain-mind is a unified system, they
are inextricably linked (ibid. p.26).
Hobson’s theory is supported by a large amount of testable,
neurophysiological, experimental evidence.
Drawing on this
evidence and the ideas of Searle, Dennett and Gelernter I believe when we
refer to ‘the mind’ we are referring to a highly complex system, which
is a combination of electrochemical-neural interactions in the organ in
the skull called the brain. The
brain interacts with the physical body, and through sensory input/output,
the environment external to it. Embodiment
with sensory input is an essential
requirement for a mind to exist. This
sensory input further allows for another essential requirement for the
creation and maintenance of mind and that is socialisation.
In a normal
human adult the unified brain-mind system has both nonconscious mental
states (memories, regulation of respiration etc.) and at times conscious
mental states. Consciousness
is simply one state of this unified brain-mind system.
Consciousness is largely controlled by the brain’s chemical
system know as the aminergic-cholinergic system.
The aminergic
system (amines) governs our waking state and the cholinergic (acetylcholine)
system governs our dreaming state. These systems are in dynamic equilibrium and neither one is
ever totally inactive. The
ratio of these chemicals can now account for many previously mysterious
states of conscious such as hypnosis, dementia and fantasy. As we approach
sleep the cholinergic chemical increases and maintains dominance whilst
asleep. As we wake up
normally, the reverse happens and the aminergic system becomes dominant.
If we are awoken suddenly we temporarily experience confusion and
disorientation because the chemical system needs a little time to
re-establish its correct ratio/balance for the respective, consciously
desired? states (Hobson, 1994. pp.14-16).
I have noted a most interesting correlation in the work of
Gelernter. He believes mental
focus moves from high to low, at the high focus end we are most alert,
logical and deal with step-by-step problem solving.
At the low focus end, that is, as we move down the spectrum we do
not think logically, our minds move easily from one unrelated subject to
another, creative solutions to problems occur at this level, ones that
have previously defied logical solution. It is at this level that
inspiration suddenly hits us. Further down the spectrum the onset of sleep
and then dreaming occurs. We must bear in mind that during REM sleep we
dream, the awareness of dreams or dream fragments, even though we are
asleep, is still part of a conscious state.
This description of mental states fits in perfectly with the action
of the aminergic-cholinergic system.
I can see that
some philosophers, though happy enough with the above, may still argue
that it does not say what consciousness actually is.
Searle helps overcome this conundrum through his belief that our
vocabulary and consequently our mode of thinking is at fault.
It is incorrect to think that a state must be either mental or
physical, Searle believes such apparent oppositions as these are false,
“Consciousness is a mental, and therefore physical property of the brain
in the sense in which liquidity is a property of systems of molecules, eg.
H2O (Searle, 1992. p.14). Further, “...consciousness qua
consciousness, qua mental, qua subjective, qua
qualitative is physical, and physical because mental (ibid. p.15).
This approach I believe is plausible in answering the elusive
question of what consciousness is.
Dennett
discussing Jaynes and Nagel, describes the chasm between inert matter and
the inwardness of a conscious being, in the example of brick and
bricklayer (Dennett, 1998. p.122). If
we accept that a brick cannot be conscious, and it is by no means a
universally accepted conclusion, many tribal societies believe inanimate
objects such as a stone do have a kind of consciousness or at least
spiritual essence. In
principle we must remain open to this because we are not stones so can
never ‘really’ know what it is like to be a stone. If we do not accept
this possibility and insist that a stone is inert and a stonemason is
conscious how can this be? Suppose
we ‘deconstruct’ both stone and stonemason.
Prior to total deconstruction we get down to molecules and atoms,
say carbon, silica, hydrogen and so on, the very same building blocks are
fundamentally present in both stone and stonemason.
Further on, we arrive at Quantum states, probabilities, particles
and waves. So how and where
from this ‘oneness’ does consciousness become an attribute of the
stonemason and not the stone? I
believe that Hobson and Searle are correct in insisting that consciousness
arises from brain-mind states. The
reason it can arise is because a functioning
system, with just the
right attributes causes it to exist.
It seems clear from scientific ‘deconstruction’ of stones that
they do not have a brain-mind system and therefore cannot be conscious in
any sense that humans are.
Franklin argues
that mind is graded not Boolean, this fuzzification of mind allows
for, some degree of mind in animals and possibly machines, though it
may be in the mechanical sense, that is, without qualia (Franklin, 1995.
p.412).
Perhaps the degree of
consciousness is proportional to the complexity of the system from which
it arises. Hence we might
imagine a consciousness complexity scale from one to one hundred, along
which; a plant may be zero, an ant two, a dog sixty, an ape seventy and a
human ninety. These figures
are of course speculative but it helps illustrate the point.
This consciousness scale has nothing to do with Gelernter’s
low-high focus of consciousness. His
model applies separately to each species which is conscious.
Rather than
providing a precise definition of consciousness, in the foregoing I have
attempted to approach the phenomena from various angles, to at least find
some things consciousness is not and others that must be present for
consciousness to arise. Two
characteristics of consciousness that are particularly relevant to this
discussion are awareness and intentionality.
Without
awareness we are not conscious, more precisely we are not conscious
‘of’ something. Awareness
may be of external events or internal brain-mind mentation, such as the
dreams of REM sleep. Various
meditational states, like dreams, have absolute minimal external stimulus
yet the individual may be consciousness of these, as an example,
consciousness of ‘nothingness’. Austin
explains much of this mysterious mental phenomena in a large body of
research work, represented in, “Zen and the Brain”
For an agent to
be considered conscious it must display Intentionality.
As Searle points out this does not mean that intentionality is
consciousness though. “Intentionality is that property of many mental
states and events by which they are directed at or about or of objects and
states of affairs in the world” (Searle, 1983. p.1). Intentionality
needs to be divided into intrinsic and ‘as-if’ intentionality for
clarification and for consideration in an artificially intelligent entity.
If a person
makes a statement such as, “I am afraid of snakes”, this is an example
of intrinsic intentionality. If
your personal computer displays a message, ‘I am afraid of snakes’
this is, as-if intentionality. Many
devices display as-if intentionality, a thermostat is a good example, but
none of these devices ‘so far’ have the presence of any mental
phenomena. If an office thermostat controlling the air conditioning had
a speaker attached and always reported when it was turning up the heating
and then one day said, “I’m not turning up the heating today because
I‘ve had it with you people treating me like a dumb wallflower!” then
it would be displaying intrinsic intentionality. Deciding between as-if
and intrinsic intentionality will be a criteria for assessing true
artificial intelligence and will not be as easy as it may seem.
For an entity to
be considered consciously aware that it exists it must possess intrinsic
intentionality, if consciousness and intentionality are features of a
unified brain-mind system we need to ask how do brain-mind states come
about, over and above their chemically controlled basis? If the brain-mind is simply an ‘information-processing’
organ, with on-off switching and access to a huge knowledge database
(memory), that is, computational power plus knowledge, we would already
have developed rudimentary artificial intelligent conscious machines.
I seems to me, in principle, this approach to AI is fundamentally
flawed.
One reason for
this is that all the information in the world is not the measure of
intelligence, one measure of intelligence is the ability of an organism to
function within its environment and survive the normal hazards of that
environment. A
walking encyclopedia will walk over a cliff, for all its knowledge of
cliffs and the effects of gravity, unless it is designed in such a fashion
that it can find the right bits of knowledge at the right time, so it can
plan its engagement with the real world (Dennett, 1998. pp.194-195).
The failure of
traditional or classical AI led to the development of the connectionist
paradigm, this included neural networks operating in parallel, similar to
the way the brain operates and accesses subsystems which, ‘do their own
thing’, at a local level. The
connectionist model allows a
system to learn and expand its program as it encounters various
situations. Whereas classic
AI (rule-and-symbol-based) is good at logic and long term planning it is
inadequate for real-time motor control and perceptual recognition (Clark,
1997. p.59).
Classic AI often
sees the computer as analogous to the human brain, little wonder scholars
such as Black argue that, “The digital computer analogy is fatally
misleading” (Black, 1991. p.3). Black
further argues that the software-hardware dichotomy is artificial,
“...software and hardware are one and the same thing in the nervous
system” (ibid.). Whilst
Black is no doubt correct in that DNA instructions are encoded right in
the cell, ‘on site’ ready to do their job, I do not agree they are
“one and the same thing”. They are instructions embedded into the molecular matrix of
particular parts of the cell.
Software
instructions are part of a machine’s memory, embedded electrically in
the matrix of the memory medium. In
all but the oldest AI programs these instructions form part of feedback
loops which modify and expand the original instructions and also rearrange
their positions and relevance in the software’s hierarchical structure.
Further to this,
it has always struck me as naive and almost absurd that AI researchers, up
until a few years ago, imagined they could create an artificially
intelligent machine and do so with the machine in isolation, simply by
increasing speed and adding more computational power.
This procedure has resulted in very powerful machines which can
outperform humans in many respects, however, this has nothing to do with
intelligence. The nurturing
period of a human infant, with the longest neoteny of any species,
together with the interaction of infant with other infants and adults is
partly the basis of human intelligence.
In this period, up to three years of age, the development of the
limbs, the structuring of neural pathways and the gradual appreciation by
the infant that it is an autonomous agent all take place.
In my opinion without an equivalent
period of infancy no machine will ever even approximate human
consciousness.
I think much
wasted discussion and programming effort has taken place because of the
limited vision of just what an AI entity would require prior to being able
to develop internal states that could give rise to consciousness.
The coming into
existence of the World Wide Web may be a great benefit for the
socialisation of AI entities, entities could be on-line for extended
periods and use the Web as a classroom for learning facts and as a place
for social interaction. I
cannot speculate on the benefits of Virtual socialisation over real
socialisation but there are futurists who regard it as equally
efficacious. Vinge in a paper presented to NASA discussed the possibility
of computer networks ‘waking up’, the Web with its millions of users
at any one time could possibly be regarded as an intelligent,
‘artificial’ entity in its own right (Vinge, 1993).
Regardless of
how an entity experiences socialisation its first requirement is
embodiment, that is, the ‘mind’ part of the entity must have some sort
of physical attributes which help locate it spatially.
An entity cannot be aware of its existence unless it has reference
to other objects which are not it.
Thelen and Smith have dome some important pioneering work in
developmental psychology which has been especially relevant in dispelling
the entrenched Cartesian notion of ‘mind’ as a separate, controlling
homunculus like thing. Known
as the Dynamic Systems approach to development of cognition and action,
this approach has proven beyond all doubt that various parts of a system,
‘do their own thing’. Literally,
the brain does not know how to do certain things nor that they have
occurred, such as some of the aspects of an infant learning to walk (Thelen
& Smith, 1994. Chap.1). These
discoveries have major implications for AI, the low-level design of the
body of an entity allows for local knowledge and control without the
burden of complex, resource hungry central executive control.
This fundamental
approach has been implemented in MIT’s, COG Project.
I was excited to come across this work as it is actually doing what
I thought were the minimum necessary basic steps in creating an
intelligent artificial entity. As
well as utilising local feedback control, the process of socialisation is
being carried out equally with hardware (body) modification and
programming evolution.
Before
describing this project in detail it is worth noting that Collins towards
the end of the eighties was insisting that socialisation and enculturation
are essential components of intelligence.
Although he did not describe the way machines must be socialised to
exhibit intelligence his work pre-empted what is happening with the Cog
project. Collins made a very
interesting point which few people seem to think about and that is,
perhaps, rather than humans trying to make machines like humans, they are
becoming more like machines themselves.
Quoting Dreyfus, “Our risk is not the advent of superintelligent
computers but of subintelligent human beings (Collins, 1990. p.190).
The Cog Project
at MIT, under the direction of Rodney Brooks, is an ongoing concern which
seeks to build “human-like artificially intelligent systems”, not
systems which master a single domain but those which can adapt to many
complex tasks in the real world in real time.
This goal has led to the rejection of many of the procedures in
classical AI and also of the assumptions about human intelligence which
are feature of this discipline.
The guiding
principle of the Cog team, is that, “...human intelligence is a direct
result of four intertwined attributes: developmental
organisation, social interaction, embodiment and physical
coupling,
and multi-modal integration (Brooks et al., 1998).
Before disusing these attributes in detail I will describe the
assumptions about intelligence which classical AI still believes, and
Brooks et al, eschew; monolithic internal models, monolithic control and
general purpose processing. Humans
have no full monolithic internal models.
When performing a copying task we do not build an internal model of
the complete scene we are attempting to copy.
Experiments have shown that, “...humans tend to only represent
what is immediately relevant from the
environment and those representations do not have full access to one
another” (ibid.). Humans have no
monolithic control.
Evidence from cognitive science whilst acknowledging control
structures finds no support for a single unitary control system.
Observation of various split brain patients suggests, “... that
there are multiple independent control systems, rather than a single
monolithic one” (ibid.). Humans are not
general purpose.
Despite the conventional, commonsense view that humans are equally good at
any tasks they attempt, experiments have shown this to be false. The way information is presented affects the ability to solve
problems quite significantly. “Humans,
often do not use subroutine-like rules for making decisions” (ibid.)
Quite often emotional rather than rational factors are the major aspects
of decision making. The work
of Damasio is seminal in this regard (Damasio, 1994.).
These three factors alone
shift significantly, the approach to designing intelligent machines and I
would add, assessing the intelligence and consciousness levels of other
natural animals. Together with the four previously mentioned attributes or
“essences of human intelligence” required in an entity it is little
wonder classical AI has not achieved its optimistic goals.
Expert Systems, such as
Weizenbaum’s famous “Eliza” program, impressive as they were, really
had little to do with true artificial intelligence (Weizenbaum, 1976).
Although numerous ‘hopefuls’ believed these systems did display
elements of true AI I do not think Weizenbaum himself ever made such
claims.
Returning now to
the essential attributes for the development of intelligence (and the
possibility) of conscious awareness I will look first at development. (a) Development
is the framework within which an infant gradually acquires more and more
complex skills. “Humans are
not born with complete reasoning systems, complete motor systems, or even
complete sensory systems” (Brookes et al. 1998). The earlier
developmental processes seem to, “...prepare and enable more advanced
forms of behavior to develop within the situated context they provide”
(ibid.). (b)
Social Interaction. “The presence of a caregiver to nurture a child as
it grows is essential. This reliance on social contact is so integrated
into our species that it is hard to imagine a completely asocial human”
(ibid.). The ABC, some three
years ago, televised some secretly obtained footage of children’s
institutions in Russia where very young children were abandoned and had
absolutely minimal social and physical contact, and certainly none with a
carer. The children were
assessed by Aid workers to be severely emotionally (and from memory)
intellectually, undeveloped. Work
with autistic children also gives clues as to the importance of social
integration, a number of scholars including Sacks, Synder and Baron-Cohen
work and investigate within this field. (c)
Embodiment. As Brookes et al.
note the most obvious and clearly overlooked aspect of human intelligence
is that it is embodied. There is a direct physical coupling between action
and perception without the need for intermediary representation.
“For an embodied system, internal representations can be
ultimately grounded in sensory-motor interactions with the world (Lakoff,
1987) (ibid.).
One reason for
it being ‘overlooked’ I believe is a religious one.
For the last two thousand years the dominant influence on Western
thinking has been Christianity. Christianity maintains the ‘flesh’ is unclean, necessary
for a time to be sure, but ultimately it is the spirit that matters.
Following in this tradition Descartes separated the body and mind,
as though the body was more or less irrelevant to the mind.
I mention this not just as an aside but because of the pervasive
influence of spiritual traditions both East and West on the psyche of
humanity. If the body is
essential to the maintenance and formation of mind and consciousness, it
raises very serious problems for such doctrines as reincarnation.
The importance
and impact of the realisation that mind and body are not separate, that
embodiment is essential to the development of anything that can be
considered ‘mind’ has not yet been realised by society at large.
The corporeal/mental (spirit-soul-mind) dichotomy is so entrenched
in our languages and culture that when it is fully realised that there is
no central controlling executive, no esoteric special ‘matter’ that
constitutes mind, it will be equivalent, in my opinion, to a Copernican
Revolution. The ramifications
of which we can only barely imagine at present. (d)
Integration. Just as no
executive controls our every function, evidence now suggests that no one
sensory input (visual, olfactory) is independent of the others.
The huge amount of information that comes from the external
environment is processed simultaneously and of course gives us our view of
the world. “Stimuli from one modality can and do influence the
perception of stimuli in another modality” (ibid.). This means any
attempt to create artificial intelligence must take this dependent phenomenon into consideration.
The Cog team’s
methodology, recognises the above attributes because they are important
aspects of human intelligence and, “...from an engineering perspective,
these themes make the problems of building human intelligence easier”
(ibid.)
Apart from the
fact that embodiment is a necessary criteria of intelligence, by giving
their Cog creations, bodies, it allows humans to interact with the robots
in a natural way. Further,
“...the effects of gravity, friction and natural human interaction are
obtained for free, without any computation” (ibid.).
One fascinating and on reflection, essential attribute of the Cog
robots, is that of eyes. The
team has recognised the vital importance of eye contact between human
infants and their carers, and later, eye contact with adults.
The robots have specially designed complex eyes which allow this
interaction and also enable the robot to visually recognise the various
people it interacts with each day.
The development of the
system is incremental, that is, the earlier learnt behaviours and so on,
“...bootstrap the later structures by providing subskills and knowledge
which can be re-used: (ibid.). Just
like a human infant the system gradually increases it understanding and
gradually becomes able to handle more and more complex problem solving
tasks. The important thing to
realise with this approach is that it is, “...in stark contrast to most
machine learning methods, where the robot learns in a usually hostile
environment, and the bias, instead of coming from the robot’s
interaction with the world, is included
by the designer” [my emphasis] (ibid.).
The Cog team’s
approach does not emphasise enculturation as much as I believe is
necessary. Socialisation is
not quite the same thing as enculturation and whilst, “Social
interaction allows humans to exploit other humans for assistance, teaching
and knowledge” (ibid.) this does not necessarily imply that culture is
being passed on per se. Culture is arguably as important in the
development of human intelligence and consciousness as are biological
factors, consequently, an intelligent entity needs to learn and be moulded
by cultural inputs so as to be able to communicate to others in that
culture. Granted, part
of this transmission of culture takes place during the normal
socialisation of humans.
One last aspect
of the search for the necessary fundamentals of intelligence and
consciousness is the notion of the Unconscious. A colleague’s chance remark, asking me just how one would
create the Unconscious, even if it was possible to create the equivalent
of the conscious mind, led me on an intense investigation of the
Unconscious. To my knowledge
this aspect of human mentation has not been discussed in the AI
literature. Similar to the
central executive controller concept being exposed as a myth so too have I
shown that the notion of the Unconscious, particularly in the Freudian
sense, is an artificial construct. The
Freudian Unconscious with its supposed sexual, libidinal repressions and
its expression, symbolically through the latent
dream content, does not exist. The
unified brain-mind has various mental states, most are nonconscious at any one time, the brain-mind may contain painful suppressed
memories but these are nothing to do with the widely accepted and almost
unchallenged existence of the Unconscious (Harle, 1999).
The removal of the Unconscious from consideration in AI research is
one further advance towards creating true artificial intelligence. In conclusion, I have attempted to present in this paper the broad issues involved in the project of creating an artificial intelligent, conscious entity. I believe this is practically and, in principle, impossible if we follow the path of classical AI, that is, computational power and huge amounts of knowledge (facts). However, if we pursue the approach developed by the Cog team, and once certain hardware constraints are overcome, especially the creation of massive parallel neural network architectures I can see no plausible argument that denies the possibility of creating an intelligent, conscious, non carbon-based entity. BIBLIOGRAPHY: Austin,
J. A. Zen and the Brain.
Cambridge, MA.: MIT press, 1998. |