What I saw as my central thesis is that propagating carefully conceived
dimensionality information along with classical "information" could greatly
improve the cognitive process, by FORCING reasonable physics WITHOUT having
to "understand" (by present concepts of what "understanding" means) physics.
Hutter was just a foil to explain my thought. Note again my comments
regarding how physicists and astronomers "understand" some processes though
"dimensional analysis" that involves NONE of the sorts of "understanding"
that you might think necessary, yet can predictably come up with the right

Are you up on the basics of dimensional analysis? The reality is that it is
quite imperfect, but is often able to yield a short list of "answers", with
the correct one being somewhere in the list. Usually, the wrong answers are
wildly wrong (they are probably computing something, but NOT what you might
be interested in), and are hence easily eliminated. I suspect that neurons
might be doing much the same, as could formulaic implementations like (most)
present AGI efforts. This might explain "natural architecture" and guide
human architectural efforts.

In short, instead of a "pot of neurons", we might instead have a pot of
dozens of types of neurons that each have their own complex rules regarding
what other types of neurons they can connect to, and how they process
information. "Architecture" might involve deciding how many of each type to
provide, and what types to put adjacent to what other types, rather than the
more detailed concept now usually thought to exist.

Thanks for helping me wring my thought out here.

On Sun, Jun 27, 2010 at 2:49 PM, Ben Goertzel <> wrote:

> Hi Steve,
> A few comments...
> 1)
> Nobody is trying to implement Hutter's AIXI design, it's a mathematical
> design intended as a "proof of principle"
> 2)
> Within Hutter's framework, one calculates the shortest program that
> explains the data, where "shortest" is measured on Turing  machine M.
> Given a sufficient number of observations, the choice of M doesn't matter
> and AIXI will eventually learn any computable reward pattern.  However,
> choosing the right M can greatly accelerate learning.  In the case of a
> physical AGI system, choosing M to incorporate the correct laws of physics
> would obviously accelerate learning considerably.
> 3)
> Many AGI designs try to incorporate prior understanding of the structure &
> properties of the physical world, in various ways.  I have a whole chapter
> on this in my forthcoming book on OpenCog....  E.g. OpenCog's design
> includes a physics-engine, which is used directly and to aid with
> inferential extrapolations...
> So I agree with most of your points, but I don't find them original except
> in phrasing ;)
> ... ben
> On Sun, Jun 27, 2010 at 2:30 PM, Steve Richfield <
>> wrote:
>> Ben, et al,
>> *I think I may finally grok the fundamental misdirection that current AGI
>> thinking has taken!
>> *This is a bit subtle, and hence subject to misunderstanding. Therefore I
>> will first attempt to explain what I see, WITHOUT so much trying to convince
>> you (or anyone) that it is necessarily correct. Once I convey my vision,
>> then let the chips fall where they may.
>> On Sun, Jun 27, 2010 at 6:35 AM, Ben Goertzel <> wrote:
>>> Hutter's AIXI for instance works [very roughly speaking] by choosing the
>>> most compact program that, based on historical data, would have yielded
>>> maximum reward
>> ... and there it is! What did I see?
>> Example applicable to the lengthy following discussion:
>> 1 - 2
>> 2 - 2
>> 3 - 2
>> 4 - 2
>> 5 - ?
>> What is "?".
>> Now, I'll tell you that the left column represents the distance along a
>> 4.5 unit long table, and the right column represents the distance above the
>> floor that you will be as your walk the length of the table. Knowing this,
>> without ANY supporting physical experience, I would guess "?" to be zero, or
>> maybe a little more if I were to step off of the table and land onto
>> something lower, like the shoes that I left there.
>> In an imaginary world where a GI boots up with a complete understanding of
>> physics, etc., we wouldn't prefer the simplest "program" at all, but rather
>> the simplest representation of the real world that is not physics/math *
>> in*consistent with our observations. All observations would be presumed
>> to be consistent with the response curves of our sensors, showing a world in
>> which Newton's laws prevail, etc. Armed with these presumptions, our
>> physics-complete AGI would look for the simplest set of *UN*observed
>> phenomena that explained the observed phenomena. This theory of a
>> physics-complete AGI seems undeniable, but of course, we are NOT born
>> physics-complete - or are we?!
>> This all comes down to the limits of representational math. At great risk
>> of hand-waving on a keyboard, I'll try to explain by pseudo-translating the
>> concepts into NN/AGI terms.
>> We all know about layering and columns in neural systems, and understand
>> Bayesian math. However, let's dig a little deeper into exactly what is being
>> represented by the "outputs" (or "terms" for died-in-the-wool AGIers). All
>> physical quantities are well known to have value, significance, and
>> dimensionality. Neurons/Terms (N/T) could easily be protein-tagged as to the
>> dimensionality that their functionality is capable of producing, so that
>> only compatible N/Ts could connect to them. However, let's dig a little
>> deeper into "dimensionality"
>> Physicists think we live in an MKS (Meters, Kilograms, Seconds) world, and
>> that all dimensionality can be reduced to MKS. For physics purposes they may
>> be right (see challenge below), but maybe for information processing
>> purposes, they are missing some important things.
>> *Challenge to MKS:* Note that some physicists and most astronomers
>> utilize "*dimensional analysis*" where they experimentally play with the
>> dimensions of observations to inductively find manipulations that would
>> yield the dimensions of unobservable quantities, e.g. the mass of a star,
>> and then run the numbers through the same manipulation to see if the results
>> at least have the right exponent. However, many/most such manipulations
>> produce nonsense, so they simply use this technique to jump from
>> observations to a list of prospective results with wildly different
>> exponents, and discard the results with the ridiculous exponents to find the
>> correct result. The frequent failures of this process indirectly
>> demonstrates that there is more to dimensionality (and hence physics) than
>> just MKS. Let's accept that, and presume that neurons must have already
>> dealt with whatever is missing from current thought.
>> Consider, there is some (hopefully finite) set of reasonable manipulations
>> that could be done to Bayesian measures, with the various competing theories
>> of recognition representing part of that set. The reasonable mathematics to
>> perform on spacial features is probably different than the reasonable
>> mathematics to perform on recognized objects, or the recognition of
>> impossible observations, the manipulation of ideas, etc. Hence, N/Ts could
>> also be tagged for this deeper level of dimensionality, so that ideas don't
>> get mixed up with spacial features, etc.
>> Note that we may not have perfected this process, and further, that this
>> process need not be perfected. Somewhere around the age of 12, many of our
>> neurons DIE. Perhaps these were just the victims of insufficiently precise
>> dimensional tagging?
>> Once things can ONLY connect up in mathematically reasonable ways, what
>> remains between a newborn and a physics-complete AGI? Obviously, the
>> physics, which can be quite different on land than in the water. Hence, the
>> physics must also be learned.
>> My point here is that if we impose a fragile requirement for mathematical
>> correctness against a developing system of physics and REJECT simplistic
>> explanations (not observations) that would violate either the mathematics or
>> the physics, then we don't end up with overly simplistic and useless
>> "programs", but rather we find more complex explanations that are physics
>> and mathematically believable.
>> we should REJECT the concept of "pattern matching" UNLESS the discovered
>> pattern is both physics and mathematically correct. In short, the next
>> number in the "2, 2, 2, 2, ?" example sequence would *obviously* (by this
>> methodology) not be "2".
>> OK, the BIG question here is whether a carefully-designed (or evolved over
>> 100 million years) system of representation can FORCE the construction of
>> systems (like us) that work this way, so that our "programs" aren't "simple"
>> at all, but rather are maximally correct?
>> Anyway, I hope you grok the question above, and agree that the search for
>> the simplest "program" (without every possible reasonable physics and math
>> constraint that can be found) may be a considerable misdirection. Once you
>> impose physics and math constraints, which could potentially be done with
>> simplistic real-world mechanisms like protein tagging in neurons, the
>> problems then shifts to finding ANY solution that fits the complex
>> constraints, rather than finding the SIMPLEST solution without such
>> constraints.
>> Once we can get past the questions, hopefully we can discuss prospective
>> answers.
>> Are we in agreement here?
>> Any thoughts?
>> Steve
>>    *agi* | Archives <>
>> <> | 
>> Modify<;>Your Subscription
>> <>
> --
> Ben Goertzel, PhD
> CEO, Novamente LLC and Biomind LLC
> CTO, Genescient Corp
> Vice Chairman, Humanity+
> Advisor, Singularity University and Singularity Institute
> External Research Professor, Xiamen University, China
> "
> “When nothing seems to help, I go look at a stonecutter hammering away at
> his rock, perhaps a hundred times without as much as a crack showing in it.
> Yet at the hundred and first blow it will split in two, and I know it was
> not that blow that did it, but all that had gone before.”
>    *agi* | Archives <>
> <> | 
> Modify<;>Your Subscription
> <>

RSS Feed:
Modify Your Subscription:
Powered by Listbox:

Reply via email to