Wednesday, March 23, 2011

The Quality of College Education

In the 1940s only 5% of americans got a college degree.  Today over 27% get a degree and over 50% of americans have taken at least some college classes. We are nolonger working with just "the best and brightest."  For this reason the quality of college education has declined. (Not to say there aren't other reasons as well. The intrusion of business into education, the decline in government funding, etc.)

Sunday, March 20, 2011

Moore's Law and Human-level Artificial Intelligence

Moore's law is often used to predict when human-level artificial intelligence will be achieved. But the
single-processor performance increase predicted by Moore's law ended in 2004 (Computer, January
2011, pg 31). This places limitations on some, but not all routes to human-level AI.  Inherently parallel
approaches (like genetic algorithms for example) can still be expected to exhibit performance growth
close to that previously seen.  The end of Moore's law for the single processor simply reduces the
number of viable approaches to (human level) AI. 

Saturday, March 19, 2011

Practical Applications of Asa H

Asa H software can be used in much the same way as neural network programs like Brainmaker, Neurosolutions, etc. It can also be used in place of some statistical routines like those in Unistat or
case-based reasoning systems like some of those described in Janet Kolodner's book Case-Based
Reasoning, Morgan Kaufmann, 1993.

Friday, March 18, 2011

Normative Intelligence

My "artificial intelligence" work might best be described as work on a normative intelligence.  Following Sutton
and Werbos I believe that a rational system "handles all of the calculations from crude inputs through to overt actions in an adaptive way so as to maximize some measure of performance over time". (Skeptic, vol. 12, # 3, pg 14, 2006)  An outline of the subsystems that might be employed to accomplish this is presented on my website (www.robert-w-jones.com) under cognitive scientist, theory of thought.  My current Asa H 2.0 software is such a system.

I have just read Tenenbaum et als paper "How to grow a mind: ..." in the current issue of Science (vol 331, pg 1279, 2011)  While I find this an excellent and useful paper I note that it lacks things like a value module and so I do not think it is a complete intelligence or "mind."  Once again I am also struck by the fact that human minds may not work in the way we would want a more rational mind to operate.

Wednesday, March 9, 2011

A Definiton of Analog Computing

Mathematical modeling studies/solves sets of mathematical equations in an attempt to predict or describe how some physical system behaves. Analog computing is just the inverse of mathematical modeling.  We track the behavior of physical systems in order to understand/solve mathematical equations.  Analog computing would not HAVE to be continuous and involve real numbers but it frequently will be.

Sunday, March 6, 2011

Setting parameters in Asa H

Asa H has a number of parameters which are not easy to set.  Th2 is a threshold value which is used to determine if two patterns are so similar that they can be considered to be the same pattern.  L is a learning rate which determines how big a step is taken during extrapolation. Th3 is a threshold that determines if a vector component is small enough to neglect.......
There are a half dozen to a dozen such parameters in Asa H and it is difficult to decide on their values.
One way to tune these parameters is the following:
The set of parameters can be treated as a vector and Asa H can be run for a period of time in a typical environment while we record the Utility gains during the run.  A second set of parameters can be employed in Asa H while it is run in the same (or very similar) environment and the Utility gain is again recorded.
Using these two vectors and utilities (or a larger set) we can then extrapolate using the Asa H extrapolation algorithm in order to improve the vector (i.e. adjust the threshold/parameter values) in an effort to improve the utility gains.  This procedure can be repeated to gradually tune the parameters.

Friday, March 4, 2011

In The Matrix

Konrad Zuse suggested that we might live inside a computer.  Because of the halting problem running a piece of software seems to be the only way to know what any given program will do.  So the problem of evil is solved, "the great programmer" can't know how his program will function until he runs it.  We who live in the program can survive death, the programmer can rerun our routine whenever he likes.

I have argued against us living in a simulation by saying that our behavior is too inefficient, our programmer would have to be incompetent.  But perhaps he is just exploring what a given set of algorithms will do and, because of the halting problem, can't know in advance how things will go.

But of course such extraordinary beliefs would require extraordinary evidence.