Dear IGI enthusiasts,
Here's a stab at an intro to a paper that I hope begins to capture the
essence of what is proposed.
I don't claim it as perfect or the final product.
What I need to know is if it speaks in a way that might lead to the change
we are looking for.
*=========================================*
*AGI Directions: towards Hybrid (H) and Synthetic (S) Forms.*
By
Dorian Aur (see previous posts)
(blame for this bit is accepted by Colin Hales
others? TBA.
1 Introduction
Here we seek to instigate a broadening of approaches to artificial general
intelligence (AGI). Be it an artificial brain the size of a worm, ant, bee,
dog or human, such an artificial intelligence is recognized here as a kind
of AGI. The original science program coined ‘artificial intelligence’ (AI)
in 1956 {refs} set sail, at the birth of computing, with a goal to create
machines that potentially have human level intelligence or better. What has
actually happened since then is the application of computers to a vast
array of technical challenges within which great successes have occurred
and are ongoing. However, in practice AI successes fell, and continue to
fall, within a now well recognized category called ‘narrow’ or
‘domain-bound’ AI. Within the atmosphere of its successes, however, the
original goal of human-level intelligence has, at least so far, evaded the
energies of a huge investment. Such has been the prevalence of this pattern
it can now be called a kind of syndrome and in recognition of that syndrome
in recent years the attainment of the original goal of human level AI has
taken on two main forms.
The first approach to human level AI one of simple assumption that by
attending to the AI ‘parts’ that the route to the AGI ‘whole’ will become
apparent or emerge naturally. This activity, now industrialised, forms the
backbone of AI investment at this present time. Its successes emerge almost
weekly now. The second approach is one of a concerted direct attack on
human-level AI. This is a recent phenomenon manifest in a comparatively
small community of investigators, with commensurate levels of investment,
who have explicitly coined the name of the goal: AGI. In doing so the
target is explicitly recognised as being of a nature deserved of an
integrated, holistic approach. This, too, is having its successes, but once
again the syndrome of narrow-AI outcomes tends to be what the practice
achieves.
Throughout all this history one thing has been invariant: The use of the
computer or more generally the use of models of intelligence as an instance
of machine intelligence. This document signals the beginning of another
approach: where the computer (model) approach is joined (to an extent to be
determined) by its natural counterpart. This new approach, for whatever
reason, is essentially untried and invisible to the AI community. It was
always an option. All we do here is get it off the shelf and dust it off as
an AGI option. This paper is a vehicle for the clear expression of an
untried approach. As such it is hoped that AI and AGI acquire a suite of
ideas and new scientific assessment techniques that will improve AI
generally as a science discipline based on a new kind of empirical testing.
Investment in the approach has been zero since day one of AI. We seek here
to make a case that if investment in this new approach was non-zero, a
cost-effective dramatic shift may occur in our understanding of the
potential kinds of machine intelligence. Specifically we seek to introduce
the concept of synthetic and hybrid AGI.
2 Computation and AGI – a perspective on practice
To understand what follows we need to carefully compare and contrast two
fundamentally different forms of computation. Formally their difference is
best captured by the words analytic computation and synthetic computation.
The first kind, analytic, is easily recognised as model-based computation.
This is where, by whatever means chosen, an abstract model is explored by
its designers. Its usefulness is inherent in what the computation tells us
upon interpretation. Within the model are representations of
characteristics that are being studied. A voltage in model may be used, for
example, to represent the actual voltage of what is being modelled. That
*representation* of something is not an *instance of* the original thing.
Recognizable forms of analytic computation include that of the analog or
digital computer (Turing machines). Its distinguishing feature is that
however the computation is carried out, its meaning is ultimately inherent
in the mental processes of a designer or in some explicit, separate
document such as software or a circuit diagram of a model. However, complex
the model is, it is best thought of as a description of something. The
description itself is the analytic form. Clearly the analytic form is
responsible for a dramatic change and technological advances in science
over decades. The computer revolution itself.
The second kind of computation, synthetic, is best understood as simply the
regularity of nature itself. Synthetic computation occurs when nature
itself is simply regarded as computation. Synthetic computation, too, may
have a designer. That is, the distinction between analytic and synthetic
computation is not held up as the distinction between ‘human-made’ and
‘naturally occurring’. Synthetic computation is when the regularity of
nature itself accepted as, or configured to be the computation. There may
be documents needed to establish the initial conditions of the
‘computation’. For example, an engineer builds and configures the initial
conditions of natural material as an automobile. The result is a synthetic
computation called ‘the automobile’ or ‘transport’. No documents are needed
to further interpret the meaning of the result of the computation. Nature
itself is the outcome of synthetic computation. Another simple example of
such computation may be seen in the concept of flight. A bird ‘computes’
those aspects of the physics of flight suited to the needs of a bird.
Humans have used those same synthetic computations (manifest in
air/fight-surface interactions) to create artificial flight. The result is
a regularity in nature accepted as a form of computation. Physically the
result is flight. That being the case, what is ‘analytic flight’? We all
recognise this: the flight simulator.
The program of future directions proposed here is one that recognises the
two different kinds of computation in the very specialized science of the
brain where the analytic/synthetic distinction can be shown to be
under-developed and potentially confused. The brain is unique in that it is
a synthetic object with a specialised role to become the natural regularity
that forms the control system of natural organisms. It embodies the
intellect of whatever creature it inhabits. The kinds of tasks such a
control system does can and have been modelled to great effect in analytic
approaches. The question is: *“What is the difference, application to the
brain, between the analytic and the synthetic approach?”* Asking that
question, and expecting a scientific answer, is what this paper is seeking.
For over half a century, approaches to creating an artificial brain have
been entirely confined to analytic forms. These analytic approaches are
explorations of models of the brain made by humans. That being the case,
then the hyper-critical issue is in understanding the conditions under
which the analytic is indistinguishable from the synthetic. If there is a
difference, then how does that difference manifest in the capability of an
AGI. For the brain, for these many decades, the synthetic half of the route
to AGI has simply been neglected for a variety of reasons. The actual
reasons for the absence of synthetic approaches to AGI is something
historians can evaluate. The practical restoration of the synthetic
approach is the goal here. The restoration of the synthetic approach is
necessary to scientifically test the difference between the analytic and
synthetic AGI. Whatever that difference is, the whole AGI enterprise has
been living within a realm of that difference for reasons that are
essentially unexplored. *Scientifically *evaluating the analytic/synthetic
difference (or the lack of it) is the goal of the proposed shift in
methodology.
In summary: The prospect of restoration of a synthetic approach to AGI is
our topic. We look at a potential change in the direction of AGI science,
and therefore the investment profile, where the analytic, the synthetic and
their hybrid are formally recognised as separate and where scientific
testing is then applied to compare and contrast their scope and
effectiveness in application to the science of the artificial brain as AGI.
In the creation of such a brain the approach can be
1.
Nil% synthetic computation (entirely analytic)
or
1.
100% synthetic computation
or
1.
H% synthetic. A hybrid form of both.
That is, the inclusion of synthetic computation to some desired level
becomes an experimental parameter. Natural brain tissue can be regarded as
naturally occurring object based on (2) synthetic computation. In
application to artificial brain tissue (AGI) so far, option (1) has been
the only approach. This has achieved all of the progress in artificial
intelligence to date. Here we suggest that the success of analytic
approaches be joined by synthetic approaches to AGI. If indeed the time has
arrived for the formal introduction of (2) synthetic AGI and (3) hybrid AGI
as viable prospects, then we need to open a discourse. What would the new
AGI science look like? What does it tell us about the scope, nature and
expectations inherent in the purely analytic approach? What does it add to
the nearly 70 year-old AGI program?
(end of section)
============================
This is offered up for discussion as the possible first part of the
document Dorian started. I have a lot more to add.
regards
Colin Hales
-------------------------------------------
AGI
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424
Modify Your Subscription:
https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657
Powered by Listbox: http://www.listbox.com
