Monday, June 21, 2010

Criticism on Novamente Cognitive Architecture, SOAR and Alikes | Критика към архитектурата Новаменте, SOAR... Курс по Универсален изкуствен разум

"Back to the Future" - an 80's classic movie and a favorite  

What's wrong with those cognitive architectures claiming to allow creation of Artificial General Intelligence (true Thinking mahichines)?

Какво не е наред в когнитивни архитектури Новаменте, СОАР и др., и защо създателите им са далеч от универсален изкуствен разум 
(истински мислещи машини, или "изкуствен интелект" както е старият лош термин, който повече им подхожда)


In brief, I've been disappointed from the papers on Ben Goertzel's Novamente  cognitive architecture (AGIRI's OpenCog seems similar), because they resemble too much SOAR-like cognitive architectures, rely on NLP, even Prolog-like "common-sense" logic.

Check out some:

http://www.novamente.net/file/AAAI04.pdf
http://goertzel.org/dynapsyc/2009/OpenCogPrime.pdf
http://goertzel.org/dynapsyc/2009/CognitiveSynergy.pdf

I have given up on a Bulgarian one, called DUAL, that uses micro-frames and a hybrid symbolic-connectionnist approach, having several levels of interaction of the entities - micro, macro... It was conceived in late 80-ies, inspired by Minsky's "Society of mind", multi-agent systems, frames, etc... It seems a bit general, but the micro-frames are to be filled manually... Also I don't see any meaningful results, achieved for 20+ years of *work*.

SOAR is "a star", it was created in late 80-ies as well, and while enhanced over time, again I haven't heard of any meaningful application of it (actually any at all), demonstrating AGI. Correct me if I'm wrong.



Architectures like this seem very elaborate at first sight, having fancy block-diagrams: they use to claim having declarative memory, episodic memory, intentional memory, attention, working memory - prides of Cognitive Science - proven experimentally to exist in human mind! They have knowledge-bases (Cyc-style are the best!), problem space, goal-states... I guess these kinds of "knowledge" are supposed to be filled *manually*, somehow like in the NLP semantic dictionary - WordNet.

I like very much the 80-ies style, I somewhat like 70-ies and 60-ies (also 50-ies),  but not really the AI-style.

Papers, journals, conferences on "Recent Advances in AI", PhDs...  And anything to impress anybody outside the conference rooms?... (Correct me if wrong)

Novamente is supposed to be better than SOAR, they claim to use probabilistic inference, trying to integrate smoothly different kinds of memory - in a really general AGI framework all those "kinds of memory" are supposed to emerge from something more general.  Novamente may work to some extent, but to me these high-level concepts, sticked together with "glue", don't look really smooth to evolve and learn well. 
 
However I partially understand guys who try to do something to seem meaningful and to sell it right now, an elaborated architecture with many blocks inside, socially acceptable and "scientifically proven" constructs might be taken more seriously than elegant abstract definitions. But:

A rule of thumb:  if an approach was feasible, for 20-some years of *work* it should had lead to a progress. To me so far it seems the same.
 
In the upcoming years, expecting:


Други ключови думи: Универсален изкуствен разум, Новаменте, когнитивни архитектури, СОАР, високо ниво, правила за извод, изкуствен интелект, Пловдив, ПУ, ФМИ, мислещи машини

No comments :