| |
|
Chapter 7: Applying Organic Design Principles to Machines is Not an Analogy But a Sound Strategy
Response to Michael Denton
Countering Michael Denton's vitalist objection that self-organizing, self-replicating, morphing, holistic forms can only be created by biological processes and that machines are necessarily deterministic and predictable, Ray Kurzweil points out that software-based, self-organizing, chaotic processes (such as genetic algorithms) can exhibit unpredictable, emergent properties and create complex original designs. Furthermore, the complexity of this "evolutionary engineering" is increasing exponentially and will match the complexity of human intelligence in a few decades, he adds.
Originally published in print June 18, 2002 in Are
We Spiritual Machines? Ray Kurzweil vs. the Critics of Strong AI
by the Discovery
Institute. Published on KurzweilAI.net on June 18, 2002.
The Bridge is Already Under Construction
Similar to Dembski, Denton points out the apparent differences
between the design principles of biological entities (e.g., people)
and those of the machines he has known. Denton eloquently describes
organisms as “self-organizing, . . . self-referential, . .
. self-replicating, . . . reciprocal, . . . self-formative, and
. . . holistic.” He then makes the unsupported leap, a leap
of faith one might say, that such organic forms can only be created
through biological processes, and that such forms are “immutable,
. . . impenetrable, and . . . fundamental” realities of existence.
I do share Denton’s “awestruck” sense of “wonderment”
at the beauty, intricacy, strangeness, and inter-relatedness of
organic systems, ranging from the “eerie other-worldly impression”
of asymmetric protein shapes to the extraordinary complexity of
higher-order organs such as the human brain. Further, I agree with
Denton that biological design represents a profound set of principles.
However, it is precisely my thesis, which neither Denton nor the
other critics represented in this book acknowledge nor respond to,
that machines (i.e., entities derivative of human directed design)
can use—and already are using—these same principles. This
has been the thrust of my own work, and in my view represents the
wave of the future. Emulating the ideas of nature is the most effective
way to harness the enormous powers that future technology will make
available.
The concept of holistic design is not an either-or category, but
rather a continuum. Biological systems are not completely holistic
nor are contemporary machines completely modular. We can identify
units of functionality in natural systems even at the molecular
level, and discernible mechanisms of action are even more evident
at the higher level of organs and brain regions. As I pointed out
in my response to Searle, the process of understanding the functionality
and information transformations performed in specific brain regions
is well under way. It is misleading to suggest that every aspect
of the human brain interacts with every other aspect, and that it
is thereby impossible to understand its methods. Lloyd Watts, for
example, has identified and modeled the transformations of auditory
information in more than two dozen small regions of the human brain.
Conversely, there are many examples of contemporary machines in
which many of the design aspects are deeply interconnected and in
which “bottom up” design is impossible. As one example
of many, General Electric uses “genetic algorithms” to
evolve the design of its jet engines as they have found it impossible
to optimize the hundreds of deeply interacting variables in any
other way.
Denton writes:
Today almost all professional biologists have adopted
the mechanistic/reductionist approach and assume that the basic
parts of an organism (like the cogs of a watch) are the primary
essential things, that a living organism (like a watch) is no more
than the sum of its parts, and that it is the parts that determine
the properties of the whole and that (like a watch) a complete description
of all the properties of an organism may be had by characterizing
its parts in isolation.
What Denton is ignoring here is the ability of complex processes
to exhibit emergent properties which go beyond “its parts in
isolation.” Denton appears to recognize this potential in nature
when he writes: “In a very real sense organic forms . . . represent
genuinely emergent realities.” However, it is hardly necessary
to resort to Denton’s “vitalistic model” to explain
emergent realities. Emergent properties derive from the power of
patterns, and there is nothing that restricts patterns and their
emergent properties to natural systems.
Denton appears to acknowledge the feasibility of emulating the
ways of nature, when he writes:
Success in engineering new organic forms from proteins
up to organisms will therefore require a completely novel approach,
a sort of designing from ‘the top down.’ Because the parts
of organic wholes only exist in the whole, organic wholes cannot
be specified bit by bit and built up from a set of relatively independent
modules; consequently the entire undivided unity must be specified
together in toto.
Here Denton provides sound advice and describes an approach to
engineering that I and other researchers use routinely in the areas
of pattern recognition, complexity (also called chaos) theory, and
self-organizing systems. Denton appears to be unaware of these methodologies
and after describing examples of bottom-up component-driven engineering
and their limitations concludes with no justification that there
is an unbridgeable chasm between the two design philosophies. The
bridge is already under construction.
How to Create Your Own “Eerie Other-Worldly” But Effective
Designs: Applied Evolution
In my book I describe how to apply the principles of evolution
to creating intelligent designs. It is an effective methodology
for problems that contain too many intricately interacting aspects
to design using the conventional modular approach. We can, for example,
create (in the computer) millions of competing designs, each with
their own “genetic” code. The genetic code for each of
these design “organisms” describes a potential solution
to the problem. Applying the genetic method, these software-based
organisms are set up to compete with each other and the most successful
are allowed to survive and to procreate. “Offspring” software
entities are created, each of which inherits the genetic code (i.e.,
the design parameters) of two parents. Mutations and other “environmental
challenges” are also introduced. After thousands of generations
of such simulated evolution, these genetic algorithms often produce
complex original designs. In my own experience with this approach,
the results produced by genetic algorithms are well described by
Denton’s description of organic molecules in the “apparent
illogic of the design and the lack of any obvious modularity or
regularity…the sheer chaos of the arrangement…[and the]
almost eerie other-worldly non-mechanical impression.”
Genetic algorithms and other top-down self-organizing design
methodologies (e.g., neural nets, Markov models) incorporate an
unpredictable element, so that the results of such systems are actually
different every time the process is run. Despite the common wisdom
that machines are deterministic and therefore predictable, there
are numerous readily available sources of randomness available to
machines. Contemporary theories of quantum mechanics postulate profound
quantum randomness at the core of existence. According to quantum
theory, what appears to be the deterministic behavior of systems
at a macro level is simply the result of overwhelming statistical
preponderancies based on enormous numbers of fundamentally unpredictable
events. Moreover, the work of Stephen Wolfram and others has demonstrated
that even a system that is in theory fully deterministic can nonetheless
produce effectively random results.
The results of genetic algorithms and similar “self-organizing”
approaches create designs which could not have been designed through
a modular component-driven approach. The “strangeness. . .
chaos, . . . the dynamic interaction” of parts to the whole
that Denton attributes only to organic structures describe very
well the qualities of the results of these human initiated chaotic
processes.
In my own work with genetic algorithms, I have examined the process
in which a genetic algorithm gradually improves a design. It accomplishes
this precisely through an incremental “all at once” approach,
making many small, distributed changes throughout the design which
progressively improve the overall fit or “power” of the
solution. A genetic algorithm does not accomplish its design achievements
through designing individual subsystems one at a time. The entire
solution emerges gradually, and unfolds from simplicity to complexity.
The solutions it produces are often asymmetric and ungainly, but
effective, just as in nature. Often, the solutions appear elegant
and even beautiful.
Denton is certainly correct that most contemporary machines are
designed using the modular approach. It is important to note that
there are certain significant engineering advantages to this traditional
approach to creating technology. For example, computers have far
more prodigious and accurate memories than humans, and can perform
certain types of transformations far more effectively than unaided
human intelligence. Most importantly, computers can share their
memories and patterns instantly. The chaotic non-modular approach
also has clear advantages which Denton well articulates, as evidenced
by the deep prodigious powers of human pattern recognition. But
it is a wholly unjustified leap to say that because of the current
(and diminishing!) limitations of human-directed technology that
biological systems are inherently, even ontologically, a world apart.
The exquisite designs of nature have benefited from a profound evolutionary
process. Our most complex genetic algorithms today incorporate genetic
codes of thousands of bits whereas biological entities such as humans
are characterized by genetic codes of billions of bits (although
it appears that as a result of massive redundancies and other inefficiencies,
only a few percent of our genome is actually utilized). However,
as is the case with all information-based technology, the complexity
of human-directed evolutionary engineering is increasing exponentially.
If we examine the rate at which the complexity of genetic algorithms
and other nature-inspired methodologies are increasing, we find
that they will match the complexity of human intelligence within
a few decades.
Denton points out we have not yet succeeded in folding proteins
in three dimensions, “even one consisting of only 100 components.”
It should be pointed out, however, that it is only in the recent
few years that we have had the tools even to visualize these three-dimensional
patterns. Moreover, modeling the interatomic forces will require
on the order of a million billion calculations per second, which
is beyond the capacity of even the largest supercomputers available
today. But computers with this capacity are expected soon. IBM’s
“Blue Gene” computer, scheduled for operation in 2005,
will have precisely this capacity, and as the name of the project
suggests, is targeted at the protein-folding task.
We have already succeeded in cutting, splicing, and rearranging
genetic codes, and harnessing nature’s own biochemical factories
to produce enzymes and other complex biological substances. It is
true that most contemporary work of this type is done in two dimensions,
but the requisite computational resources to visualize and model
the far more complex three-dimensional patterns found in nature
is not far from realization.
In discussing the prospects for solving the protein-folding problem
with Denton himself, he acknowledged that the problem would eventually
be solved, estimating that it was perhaps a decade away. The fact
that a certain technical feat has not yet been accomplished is not
a strong argument that it never will.
Denton writes:
From knowledge of the genes of an organism it is impossible
to predict the encoded organic forms. Neither the properties nor
structure of individual proteins nor those of any higher order forms—such
as ribosomes and whole cells—can be inferred even from the
most exhaustive analysis of the genes and their primary products,
linear sequences of amino acids.
Although Denton’s observation above is essentially correct,
this only points out that the genome is only part of the overall
system. The DNA code is not the whole story, and the rest of the
molecular support system is needed for the system to work and for
it to be understood.
I should also point out that my thesis on recreating the massively
parallel, digitally controlled analog, hologram-like, self-organizing
and chaotic processes of the human brain does not require us to
fold proteins. There are dozens of contemporary projects which have
succeeded in creating detailed recreations of neurological systems,
including neural implants which successfully function inside people’s
brains, without folding any proteins. However, I understand Denton’s
argument about proteins to be an essay on the holistic ways of nature.
But as I have pointed out, there are no essential barriers to our
emulating these ways in our technology, and we are already well
down this path.
Denton writes:
To begin with, every living system replicates itself,
yet no machine possesses this capacity even to the slightest degree….Living
things possess the ability to change themselves from one form into
another…The ability of living things to replicate themselves
and change their form and structure are truly remarkable abilities.
To grasp just how fantastic they are and just how far they transcend
anything in the realm of the mechanical, imagine our artifacts endowed
with the ability to copy themselves and—to borrow a term from
science fiction—“morph” themselves into different
forms.
First of all, we do have a new form of self-replicating entity
that is human-made, and which did not exist a short while ago, namely
the computer (or software) virus. Just as biological self-replicating
entities require a medium in which to reproduce, viruses require
the medium of computers and the network of networks known as the
Internet. As far as changing form is concerned, some of the more
advanced and recent software viruses demonstrate this characteristic.
Moreover, morphing form is precisely what happens in the case of
the reproducing designs created by genetic algorithms. Whereas most
software viruses reproduce asexually, the form of self-replication
harnessed in most genetic algorithms is sexual (i.e., utilizing
two “parents” such that each offspring inherits a portion
of its genetic code from each parent). If the conditions are right,
these evolving software artifacts do morph themselves into different
forms, indeed into increasingly complex forms that provide increasingly
greater power in solving nontrivial problems. And lest anyone think
that there is an inherent difference between these evolving software
entities and actual physical entities, it should be pointed out
that software entities created through genetic algorithms often
do represent the designs of physical entities such as engines or
even of robots, as recently demonstrated by scientists at Tufts.
Conversely, biological physical entities such as humans are also
characterized by the data contained in their genetic codes.
Nanobots, which I described in the first chapter of this book,
will also provide the ability to create morphing structures in the
physical world. J.D. Storrs, for example, has provided designs of
special nanobots, which he calls “foglets,” which will
eventually be capable of organizing and reorganizing themselves
into any type of physical structure, thereby bringing the morphing
qualities of virtual reality into real reality.
On Consciousness and the Thinking Ability of Humans
Denton writes:
Finally I think it would be acknowledged by even ardent
advocates of strong AI like Kurzweil, Dennett and Hofstadter that
no machine has been built to date which exhibits consciousness and
can equal the thinking ability of humans. Kurzweil himself concedes
this much in his book. As he confesses: “Machines today are
still a million times simpler than the human brain. . . . Of course
Kurzweil believes, along with the other advocates of strong AI that
sometime in the next century computers capable of carrying out 20
million billion calculations per second (the capacity of the human
brain) will be achieved and indeed surpassed. And in keeping with
the mechanistic assumption that organic systems are essentially
the same as machines then of course such machines will equal or
surpass the intelligence of man. . . . Although the mechanistic
faith in the possibility of strong AI still runs strong among researchers
in this field, Kurzweil being no exception, there is no doubt that
no one has manufactured anything that exhibits intelligence remotely
resembling that of man.
First of all, my positions are neither concessions nor confessions.
Our technology today is essentially where I had expected it to be
by this time when I wrote a book (The Age of Intelligent Machines)
describing the law of accelerating returns in the 1980s. Once again,
Denton’s accurate observation about the limitations of today’s
machines is not a compelling argument on inherent restrictions that
can never be overcome. Denton himself acknowledges the quickening
pace of technology that is moving “at an ever-accelerating
rate one technological advance [following] another.”
Denton is also oversimplifying my argument in the same way that
Searle does. It is not my position that once we have computers with
a computing capacity comparable to that of the human brain, that
“of course such machines will equal or surpass the intelligence
of man.” I state explicitly in the first chapter of this book
and in many different ways in my book The Age of Spiritual Machines
that “this level of processing power is a necessary but not
sufficient condition for achieving human-level intelligence in a
machine.” The bulk of my thesis addresses the issue of how
the combined power of exponentially increasing computation, communication,
miniaturization, brain scanning, and other accelerating technology
capabilities, will enable us to reverse engineer, that is to understand,
and then to recreate in other forms, the methods underlying human
intelligence.
Finally, Denton appears to be equating the issue of “exhibit[ing]
consciousness” with the issue of “equal[ing] the
thinking ability of humans.” Without repeating the arguments
I presented in both the first chapter of this book and in my response
to Searle, I will say that these issues are quite distinct. The
latter issue represents a salient goal of objective capability,
whereas the former issue represents the essence of subjective
experience.
In summary, Denton is far too quick to conclude that complex systems
of matter and energy in the physical world are incapable of exhibiting
the “emergent . . . vital characteristics of organisms such
as self-replication, “morphing,” self-regeneration, self-assembly
and the holistic order of biological design,” and that, therefore,
“organisms and machines belong to different categories of being.”
Dembski and Denton share the same limited view of machines as entities
that can only be designed and constructed in a modular way. We can
build (and already are building) “machines” that have
powers far greater than the sum of their parts by combining the
chaotic self-organizing design principles of the natural world with
the accelerating powers of our human-initiated technology. The ultimate
result will be a formidable combination indeed.
Copyright ' 2002 by the Discovery
Institute. Used with permission.
| | Join the discussion about this article on Mind·X! 
| |
