The Holy Grail of AI - an Impossibility

 The following is an insurmountable problem with regard to designing an

 

intelligent machine:

 

 We as human beings have various senses. We experience phenomena through our

 

senses. The axiomatic concepts that we understand are the product of our "experience",

 

e.g. Distance, Point, Weight, Color and Sound.

 

The most meaningful way that we can "know" something is to experience it. Experience

 

is our foundational method of knowing. 

 

Unfortunately, a machine lacks senses. It has only one sense: "on" or “off”

 

and therefore, the vast complexity of knowledge implicit in man's

 

conceptual content is something that a machine can not possibly  "understand",

 

because the machine has not "experienced" the phenomenon in question.

 

 "Everything" would be "abstract"  (i.e. not directly tied to reality)  to

 

the machine and hence not truly knowable. Therefore, it is impossible to design

 

a machine capable of human-like thought.  For us, Consciousness, (i.e. Awareness)

 

or Experience are  sensory phenomena and therefore, our "knowledge" is sensory

 

based. Our knowledge takes place at a higher level than discrete physical phenomena.

 

Rule based, domain specific systems are the best we can hope for in AI.

Ted,

 

Your right about my a priori assumption of course.

 

The qualia issue however is central to my position with regard to the "level" of conceptual complexity that a given machine could achieve. If a space alien had different sensory systems then the complexity of its concepts could be comparable or greater than ours, although probably quite different in my respects.

 

My position is that a machine of whatever kind couldn't achieve a  "comparable" level of conceptual complexity unless it could treat qualia issues as axiomatic, for instance, the way humans do. For a machine to "understand" and then integrate various qualia its sensory inputs would have to convert that information to a form manageable to itself. The conversion process itself would cause a loss of the richness of that information. If from birth Helen Keller lacked mobility and the sense of touch as well as her blindness and deafness how would she have grasped such fundamental concepts as "mass" and "distance" and then what would that have done to her ability to understand the world.

 

As for my personal information, I have an undergraduate degree in the liberal arts as of 1971. That is as far as I could bare to put up with Academe. I've been a software consultant and have engaged in various entrepreneurial pursuits from would be movie producer to futures speculator and many points in between. With a background in Psychology and Philosophy I found myself attempting to design a general purpose Inference Engine several decades ago. I smacked into a brick wall of a design problem the fruits of which you have been reading.

 

Regards,

Howard  

Hello again Stephen,

 

I've re-read your last post. It seems we're in agreement as regards a general purpose inference engine. Am I wrong?

 

As regards the distinction between the real and the un-real, we humans have trouble with that ourselves, so I'm not sure what your getting at.

 

I don't think "life" is necessary in order to solve problems. The key issue is the term "understand", which you used yourself. The way we use that term implies consciousness. A general purpose inference engine would likewise need to deal with that concept as fully as we do and I don't believe that is possible. By the way, a non-living machine does face problems such as the need for energy, or to be flippant, oil.

 

Does life imply intelligence? What of a virus? I suppose if we ever succeed in creating a living machine then ultimately a fully intelligent one would just be an engineering problem.

 

With regard to the law of identity issue, you hit the nail on the head and nicely put I might add.

 

I don't deny that AI research can help us understand our own problem solving methodologies. I don't think it's usually the most cost effective way however.

 

Your description below captures the approach that was adapted for AI research and I think correctly so given the inherent limitations of a system that lacks consciousness "as we know it".

 

You wrote: 

 

 

Intelligence, as we think of it, is the capacity to solve problems. The primary purpose to which we put our intelligence is to solve problems. But this formulation already implies conceptual isolation and integration. These goal directed abstraction processes and their on the fly redirection with regard to multi-facetted data already imply awareness and self awareness. To the extent problem resolution is going to be useful to us its going to be so complex and multi-sensory based and abstract as to imply consciousness. But this is central to my initial contention. As for proof of this I don't know that I can add more to what I've already said.

 

Her's a question: can their be consciousness without any sensory input? To be conscious is to be aware of something. Now there's a tautology. Developmentally speaking, if a human were born without any of its sensory systems functioning even in the slightest way would it ever be conscious of anything including its own mind? I think not.

 

You wrote:

"Physical symbol systems solve problems by searching. A solution (if there is one) is a path to a certain symbol structure; attainment of that structure is the goal. In problem solving systems, there is a test by which the system will be able to recognize that the goal state has been reached, a set of states (the search space) among which is the goal state, and a procedure for generating sequentially each candidate state for testing until the goal state is reached.

For many problems, the search space is enormous. For such problems, search of the space needs to proceed by a not entirely uninformed, selective sequence of candidate states. A somewhat informed, selective search is a heuristically guided search. A heuristic search may not find the goal state in the allotted time, but its chances of finding the goal state in that time are better than random selection of candidate states. A physical symbol system exhibits intelligence to the extent that it can generate sequences of candidate states in an order such that the goal state has "a high likelihood of appearing early" (Newell and Simon 1976). This was a milestone in the scientific understanding of intelligence. The intelligence displayed by the system is made possible by putting into the generator procedure heuristic knowledge about, and sensitivity to, probable structure of the search space so as to guide the generation of sequences of candidate states, thence to guide the search for the goal state."

 

Does this approach work for "all" non-predefined problem resolutions or just for certain limited goal states?

 

As for situated cognition, I checked out your supplied link. I still need a clear definition. Would the following suffice? Situated cognition is a theory of mind that views the nature of mind as encompassing the world external to and including the biological body that is a necessary condition for the mind's existence.

Hello Ed,

 

I checked out the link you gave me. Thanks.

 

The emergent properties that would be necessary for true AI at the human level of efficacy would be analogous to some combination of vision, touch, hearing etc., (maybe just touch alone). This in combination with the ability to self direct. Self direction has already been achieved, at least partially, as in the ability to recognize whether a goal has been achieved. It's at the foundational level of qualia and its implication for conceptual complexity that I have fundamental doubts.

 

Howard