| 
AI Ramblings 
Monday 07 February 2011  News from the AGI War Front
Thru the hustle and bustle of overoptimistic and/or scaremongering blather about AGI and the Singularity a more reasonable notion is slowly emerging.
There will not be any Singularity (or it already happened) and AGI may not be what you think.
Of special interest are :
The Undesigned Brain is Hard to Copy by Kyle Munkittrick.
A New Direction in AI Research by Monica Anderson.
In response to some comment at vetta.org I gave my own summary which I reproduce below.
Computation only occurs over an already encoded model of reality , that is, the bits shuffled around by the computation(s) have a meaning which relates to some ontology of concepts and objects in/of the model.
It is the buildup of this ontology which is the problematic step which is glossed over by current AI research.
This does NOT refers to some would be “mystical properties” of the brain or neurons but to the feature extraction phase of any problem solving method: What are the relevant features you have to sift out of the raw mass of data inputs to solve the problem at hand?
E.g. before you set to solve a chess problem you have to “know” what the pieces and the chessboard are and what are the rules.
In chess this is a given which comes with the problem statement, in any AGI problem, beside possibly the goal, the relevant concepts are NOT part of the problem statement (score best in “this environment”) they have to be created.
Of course this is ALSO a matter of computation(s) but WHICH kind of computations?
Best summarized by Nick Szabo:
The most important relevant distinction between the evolved brain and designed computers is abstraction layers.
Being “turing complete” buys you no edge in this game...
Update: IBM Watson jeopardy player defeating Jennings and Rutter triggered another frenzy of idiotic comments and undeserved awe about "Our New Computer Overlords" , actually it is not so much an advance in AI than an improvement on technology in capacity, speed and accuracy.
See discussion at Dick Lipton and John Langford blogs.
Technical overwiew Building Watson An Overview of the DeepQA Project (via Carson Chow).
Submitted by KevembuanggaKevembuanggaonMonday 07 February 2011 - 19:10:14   comment: 2 Wednesday 18 November 2009 Objects as epistemological artifactsTo follow on a few arguments I had at some blogs like Gödel’s Lost Letter, vetta project, Machine Learning (Theory) and The n-Category Café I think I need to clarify the matter about my personnal stance on "objects" and ontologies, this also for my own benefit.
I am strongly anti-platonist because all evidence points to the irrelevance of metaphysical postures for the actual practice in mathematics and engineering.
There is no need to question and look and check for the actual existence of objects and concepts (which are also objects, supposedly abstract and immaterial) "in reality" because objects and concepts are NOT part of reality but part of our representation of reality (whatever one's view of reality is).
We shoudln't fear to be lacking of any sort of objects as representations (including pink unicorns) because as shown by the "existence" of the Rado Graph any fancy (countable) structure can be found.
The only trouble we can have is improper use of language (mathematical or otherwise) in our attempts to pin down "an object", that is a lack of definedness, mistakenly expecting that a sentence (or any sort of syntactical construct) actually points (denotes) an object (one and the same) in a consistent manner.
As I hinted the only purpose of objects is to serve as carriers of properties in our discourse.
Think about it for a moment, how could we organize any display of information if it were not possible to refer to the "same thing" at two or more distinct places in a discourse of any kind?
This is why objects must be intemporal (eternal), we already have enough trouble with the shifting of meanings due to the fuzziness of actual communication practices without having an indeterminacy "in principles".
This leads the Platonists to believe in the existence of abstract objects as some ghostly duplicates (non-physical and non-mental) of material objects, this is only a clumsy projection of folk intuition.
Objects are referents not "things" in the lay meaning of the word.
Furthermore ANY POSSIBLE OBJECT potentially exists, any piece carved out of the Rado Graph can serve as a referent, a label, a pointer in a discourse, being both recognizably distinct and unique (up to isomorphism).
Trying to sort out which objects "exist" like the Platonists do is devoid of any meaning, because all do exist.
Which doesn't mean that we should have an interest in every of them or will ever meet them all.
As René Thom said "Truth is not limited by falsity, but by insignificance".
Therefore what's the point with ontologies?
Ontologies are not just lists of "existing objects" they necessarily involve some language with wich they define the objects and their relationships.
And this is the valuable part of ontologies, they establish the basis for some discourse.
They also enforce a somewhat arbitrary partition of the "reality" they aim to be about, what is known in linguistic as the Sapir-Whorf Hypothesis.
However there is no one true right ontology it all depends on the problems at hand and even for the same one problem (or class of problems) there are many possible ways to "ontologize" it, like foundational problems in mathematics.
What makes the difference is the convenience of the ontology relative to the questions sought for.
This is why quasi-religious haggling about the "right way" to talk or think about this or that is pretty pointless.
Yet, shifting our perspective (swapping/altering ontologies) is something we do so naturally and with so much ease (if not rigor  ) that we forget how critical it is for our thinking process.
The many different proofs of any given theorem require such "translations" between sligthly different perspectives, at least with respect to the lemmas used in the proof which, though may be defined inside a same general framework, are not necessarily related in an obligatory manner. And, may I remind you, lemmas ARE OBJECTS ON THEIR OWN, they have names, they can be recognized and beside terminology quarrels they have unicity.
It is this ability to build objects tailored to a purpose which is the key to our ability to deal "intelligently" with the world, what Barwise & Etchemendy call heterogeneous reasoning.
As far as I know The Mutilated Checkerboard problem is still NOT solved in AI except by brute force because it requires (so called...) human creativity in choosing a clever approach, i.e. shifting the perspective.
This is why I object to the simplistic view of Marcus Hutter and als, that AI is about sequence prediction.
I also deem all "foundational quarrels" in mathematics entirely irrelevant.
What needs to be done is to figure out what we are exactly doing when we shift perspectives and juggle wit h ontologies, because WE DO IT (successfully...)
Submitted by Kevembuangga KevembuanggaonWednesday 18 November 2009 - 18:20:43 comment: 6 Thursday 06 November 2008 Almost right... I found a paper by Diederik Aerts about composite concepts recognition which in spite of the mouthful of the title (General Quantum Modeling of Combining Concepts: A Quantum Field Model in Fock Space) hints at a very simple and always overlooked fact: We do not use any "obvious" logic in our actual appraisal of category membership.
Following several previous papers of his own about the use of quantum logic instead of classical (boolean) logic this one sum up the whole of his ideas on the matter and demonstrates a stunning adequacy to experimental results from cognitive psychologist James Hampton.
To quickly highlight the main point, it amounts to have the concepts somehow warp their "expected" combination (conjunctive or disjunctive) depending on the instance member to be tested for.
To this purpose Diederik Aerts deploys the full gear of Quantum Field Theory with excellent results, however...
Beside being quite an overkill this monstrous apparatus is still in need of a little help from the experimenter in the form of "appropriately choosen" quantum angles which parametrize the sought for quantum interferences.
With such finely hand tuned values Aerts quantum model of concepts combination gets an astounding zero discrepancy with respect to almost all experimental measurements of human responses.
Not only is it highly suspicious to have a "perfect match" with psychological experiments necessarily subject to measurement errors but, alas, a few unruly sample cases refuse to bow to the quantum model dictums no matter the choice of the "quantum angles", what a pity!
This otherwise outstanding piece of work suffers from a common disease, trying to forcefully shoehorn the reality into a whimsical model instead of trying to unravel more cogent causes for the observations.
From both the succesful cases (even if a bit cheated) and the failing ones it should be obvious that: - Whenever concepts are combined in a composite query they interact in intricate "non classical" ways.
- Quantum logic as such isn't the perfect match to model this interaction.
Why the heck would it have to be?
Like the Platonic Solids were to match the planets orbits?
That was by itself an excellent enlightening exercise in new concepts formation, but now would somebody please give us the actual model of concepts interactions which show up in Hampton's experiments?
 Submitted by Kevembuangga [/html] KevembuanggaonThursday 06 November 2008 - 19:42:35 comment: 4 Thursday 15 May 2008 The smell of Strong AI In my decades long survey of the field I had totally overlooked the works of Jerry Hobbs, yet he seems on the right path, still a bit entangled in "ontology problems" may be and having too much interest in the silly Semantic Web (I hope it's only a matter of budgets...) but I think he is definitely tackling the right questions at the meeting point of Language and Logic.
 Submitted by Kevembuangga
KevembuanggaonThursday 15 May 2008 - 19:04:44 comment: 1 Sunday 06 January 2008 Roger Schank defectsIn a recent answer to The Edge Annual Question Roger Schank said:
When reporters interviewed me in the 70's and 80's about the possibilities for Artificial Intelligence I would always say that we would have machines that are as smart as we are within my lifetime. It seemed a safe answer since no one could ever tell me I was wrong. But I no longer believe that will happen. Of course, GOFAI won't do it!
The wrong headed directions promoted by the biggest stars of AI have done much, much harm to AI, even leading (in 1984 already!) to a panel discussion at AAAI-84: The Dark Ages of AI where Schank himself acknowledged the problem, even later elaborating on his own view of the difficulties he said:
It is not that AI needs definition, it is more that AI needs substance Wrong!
AI does need a definition!
No wonder that this didn't spur any improvment in the field, the culprit was the obsessive focus on logic that he now recognize in the Edge answer :
Early AI workers sought out intelligent behaviors to focus on, like chess or problem solving, and tried to build machines that could equal human beings in those same endeavors. While this was an understandable approach it was, in retrospect, wrong-headed.
Yet, he doesn't really grasp why this was wrong-headed, it's not because "Chess playing is not really a typical intelligent human activity".
He hints at the real cause ("How can we imitate what humans are doing when humans don't know what they are doing when they do it?") but doesn't quite come thru to the full conclusion:
We need to know what we want to do, no just tinker around with "promising" research projects.
The various current crazes about natural language processing, robotics and even statistical learning will be as deceptive as GOFAI, may be bringing a few high-tech gimmicks on the side but not of more decisive import than expert-systems which were the main outcome of GOFAI.
The common trap in which all AI researchers seem to fall is that they are always working on already human-elaborated concepts and whatever "good tricks" they find for processing and "massaging" those concepts more efficiently they are oblivious to the fact that they made up the concepts at start not the computer.
The most basic question therefore seems to be what do we put in a concept?
When and how do we come up with the words we use to talk and think, apple, water, red, line, cosine, mountain, etc...
In the rest of the Edge answer Roger Schank basically give up.
Well, good riddance sir...
One less nearly useless budget eater!
 Submitted by Kevembuangga KevembuanggaonSunday 06 January 2008 - 17:48:16 comment: 2
News Categories
| |
[news] Almost right...
