Wednesday, June 29, 2011

(How) Can Asa H (or other AIs) Exhibit Consciousness?

Scientists have proposed perhaps a hundred different definitions of intelligence.  Asa H (Trans. Kan. Acad. Sci., vol. 109, # 3/4, pg 159, 2006) satisfies most all of these.  But the question of consciousness is a more difficult issue.
There are a number of different theories of consciousness (Some Theories of Consciousness, Ann. Mtg. Kan. Aca. Sci., Hutchinson, April 2000, R. Jones):
#1. There is no consciousness (behaviorism.
#2. Consciousness exists but plays no useful role (Roger Carpenter).
#3. The brain acts like layer after layer of feature detectors (Comm ACM, Nov. 1990, pg 63, fig. 8) starting from things like edge detectors and leading into something like grandmother cells (though these can be distributed).  You are conscious of your grandmother when this cell (or set of cells) is active.
#4. Feedback is the key to consciousness.  As in Elman neural networks you can see your own thoughts/internal signals/internal state as feedback inputs from the hidden layers.
#5. Consciousness is the contents of the global workspace, the blackboard of a blackboard architecture.
#6. Spreading activation in a semantic network.  What are conscious are the active nodes.
#7. Message aware control system (Scientific Approaches to Consciousness, Schneider and Pimm-Smith, pg. 65, Cohen and Schoolers eds., Lawrence Ehrlbaum, 1997)
#8. Metaprocessing.  Modules watching modules.
#9. Self-consciousness is your model of you, which is a part of your model of the world.
#10. Consciousness is a serial algorithm running on parallel hardware (Dennett, Consciousness Explained, Little Brown). Leads to feedback.
#11. Consciousness is the contents of the various modules' buffers (J. R. Anderson, How can the human mind occur in the physical universe?, Oxford U. Press, 2007, pg 243)
#12. Consciousness is emergent.
#13. There are multiple (various kinds of ) consciousness, many of the above.
If theories 1, 2, or 3 are correct there is nothing we need to do with Asa H.  A recurrent network like an Elman network can be unrolled and equated with a purely feedforward network.  This could be equated with a multiple layer Asa H network.  Alternatively, feedback links can be added to Asa H. This might take care of theory 4. A blackboard is a memory with multiple access (and, possibly, access control logic). If the messages between the Asa H layers are equated to blackboards this would take care of theory 5.  Asa H is a kind of semantic network as in theory 6.  As in theory 8 Asa H has certain modules (extrapolators, deduction system, etc.) watching other modules and upper layers watching the contents of the lower layers.
Are these the right kind of modules?  Are they watching the right things? Asa H can form grandmother cells.  One of these could be a "me" cell (as would develop with a robot seeing itself in a mirror for instance).  This takes care of theory 9. We would like to run Asa H on parallel hardware, theory 10.  Asa H has the kind of buffers needed for theory 11.  Do we have the right modules? Theory 12 usually assumes that emergence occurs as the size of the semantic system increases.  We are constantly trying to scale Asa H up.
Theory 7 is not a clear match with Asa H.
Asa H (and other AIs) may be conscious depending on what is the correct theory of consciousness.

Thursday, June 23, 2011

ebooks

Of the last 14 books I bought 3 were available in electronic form. (Actually, 1 was kindle and 2 were nook and I only own a kindle reader.)  Also, the price of an ebook is usually a substantial fraction of the price of the book in paper.  In fact, the price of an ebook is sometimes greater than the price of the book in paper (a current example is Minsky's "The Emotion Machine").

Sunday, June 19, 2011

Creativity experiments with Asa H

I have argued that our concepts and our science is not unique (my blog of 29 August 2010 and of 31 December 2010 , my poster at the American Physical Society meeting Bull. Am. Phys. Soc., vol. 55, #1, 2010 and my website http://www.robert-w-jones.com/). Asa H learns its own concepts at each layer in its hierarchy.  It will be interesting to see what concepts it forms, where they resemble our own, and where they are unique. By adjusting thresholds so as to require finer scale clustering one can produce more and more precise concepts.

Sunday, June 5, 2011

Themes

In my blogs of 16 Sept. 2010 I gave advice on how to begin research and then build up a research program. This is a continuation of that advice.

Some researchers are "polymathic," their research is diversified.  A good example would be Leo Szilard who did work in:
thermodynamics
X-rays in crystals
nuclear physics
metallurgy and engineering (his patent of the nuclear reactor with Fermi and his weapons work)
bacteriology
biochemistry

There are disadvantages to this style of work.  More time must be spent on "getting up to speed" in each new research area.  One must learn different tools, methods, and basic knowledge and get to know and be known by different sets of colleagues.  This impacts what journals you can hope to publish in and what funding you can hope to obtain.

Certainly nowadays interdisiplinary work is very important but for many researchers it is better to stick to some research theme.  One theme I followed in my fusion energy work was open system end plugging.  A more popular fusion energy theme is tokamak studies.

Sticking to a research theme has various advantages.  It is easier to keep a steady source of funding.  Once learned, you stick to a common set of tools, methods, and basic knowledge.  You become known in the community and by the journals you need access to in order to publish your results.

Friday, June 3, 2011

More Experiments With Asa H

I have been working on the following abstract for a conference early next year:

Our recently developed "Asa H" software architecture (KAS Transactions 109 (3/4): 159-167) consists of a hierarchical memory assembled out of clustering modules and feature extractors. Various experiments have been performed with Asa H 2.0:
1. prefer extrapolation from real recorded patterns over extrapolation from synthetic cases
2. record signal input only when it changes by several standard deviations
3. include the number of times a pattern has been seen as a contribution to this pattern's utility
4. weight a pattern's feature more if the feature's standard deviation is smaller
5. search for short sequences common to longer patterns which have high utilities
6. search only until a "close enough" match is found
7. if a component of a case vector varies less (as measured by its standard deviation) then value it (weight it) more
8. as a postprocessor Asa outputs can be made the reference input commands to adaptive approximation based controllers
9. after some number of tries if we can't improve all components of a vector utility then approximate with a scalar utility
10. store and update a mean and standard deviation for time warpage and spatial dilation.  Use these as components in the subsequent degree of matching
11. longer memory for very low utility cases (so we can avoid them)
12. prefer pattern changes that involve agent output change
13. try randomness detection as a filter at the input and at other levels in the network hierarchy
14. compression by blending; if V1 and V2 are similar enough replace them with their vector average

Wednesday, June 1, 2011

(One sort of*) Attention

In Asa H (Trans. Kansas Acad. Sci., vol. 109, pg 159, 2006) a similarity measure (usually a vector dot product) is used to decide if an input vector, IN, falls into some category, Ini. If the dot product of IN with Ini > ThA (some threshold value) then IN is a member of category i. 

One can add a second threshold, ThB < ThA and if IN dot Ini < ThB then IN is not a member of category i.

If, however, ThB < IN dot Ini < ThA then we can look more closely at the degree of match between IN and Ini.  For instance, we can tune the time warping or spatial transformations like scaling, shifting, rotating, etc. to see if a better match is possible (>ThA).

*Various sorts of attention have been employed in Asa H.  As the input comes in as a function of time we may stay with the currently active case/category, i , so long as the match remains strong enough (similarity measure exceeds some threshold TH).  We can avoid search until the match drops below TH.  This is a sort of attention (to category i) present in Asa H 2.0 lite (see my blog of 10 Feb. 2011)