Soft State Magic Forums - Whatever Happened to AI? Article by Doug Lenat

Doug Lenat, one of the leading A.I. people still active in that field, and now president of Cycorp has a very insightful article in the Summer 2008 edition of The AI Magazine entitled "The Voice of the Turtle: Whatever Happened to AI?". I am sure Doug refers to himself as the turtle for his tenacious work on bringing to reality common sense reasoning for language queries. He lists 12 reasons why AI has failed to really bring about machine intelligence. His reasons also apply to artificial neural networks.

12. The Media and the Arts. Here Lenat refers to the hype cycles that beset A.I. in which funding first becomes too easy only to dry up completely as the promises made by most researchers are not met. This leads to sub-optimal funding decisions.

11. Smurfitis. As a result of competing for funding researchers in academia tout their ideas as being something new and give their ideas new names when in reality their ideas are not new. This leads to confusion in terminology.

10. Cognitive Science. Here Lenat argues that trying to mimic the brain is a waste of time. Instead one should be trying to understand the brain's underlying principles so one can adapt those principles to human technology and even improve on them in certain situations. The goal is to find the right balance between learning and pre-design of the system.

He did not mention this example but the recent simulation mimicing a cortical column down to the synaptic level using IBM's Blue Brain supercomputer falls under this failure. In Blue Brain's case the researchers are trying to convince everyone that their production of low frequency waves like those found in the brain is a big success even though the whole multi-million dollar simulation does not do anything else.

Lenat's has a bias more towards pre-design than myself by indicating that he thinks the symbolic algorithmic AI level is appropriate. In contrast I think parallel simulations at the circuit level based upon Soft State Automata theory is the best path to ultimate success. I do not see his approach ever working in an autonomous robot because it takes out direct emotional and spatial associations. In addition it cannot handle uncertainty well and not being a parallel circuit it relies on search algorithms instead of cue recall.

Another time waster in most cases is building robots. Here I must agree as well because lots of researchers have built robots without having any idea about how to make them intelligent and now those robots just sit idle. But that should soon change.

9. Physics Envy. Another time waster according to Lenat is the attempt to develop general "laws" of intelligence" in the sense of formal logic proof theorems. Again I must agree. These "theories" try to model themselves after Euclid's geometry by assuming all knowledge can be deduced from a few fundamental theorems. This is a deductive only process. In fact Lenat's own initial work involving theorem proving which chains facts together derives from this approach so his opposition to this is significant.

Yet knowledge is not only acquired from deduction but also from induction which assumes some causal connection via an association. Logical "proofs" using induction make use of the IMPLICATION operation which assumes a statement is true unless contradicted. Not surprisingly the proofs of Kurt Godel show that these formal theories can be inconsistent. Yet underlying causal flows do exist as shown by Soft State Automata and without them reality will not exist, the mechanisms of the universe and the brain would not exist. The goal is to create inductive theories that capture the essence of the causal flows within some domain. These theories are true up to their limit of explanatory power.

Soft State Automata is an inductive theory and as such it is is more in the style of foundational physical theories like Newton's Theory of Gravity or Communication Theory as developed by Claude Shannon in which novel concepts are proposed and justified. Unfortunately deductive theories rule in computer science at the moment with many of their true believers controlling publications and funding sources.

8. Jerks in Funding. Again this is related to faddish funding cycles. At the present time brain intelligence research now has little funding. What funding that has existed recently is very narrowly targeted with little risk and little chance of progress in the bigger issues. In addition the research programs bring in smaller amounts of money for shorter periods of time. Because of this limited funding researchers in academia must now spend most of their time angling to get funding. Many are getting out of the field altogether.

7. Academic Inbreeding. Lenat bemoans the increasing specialization and tribalism in academia. Each tribe (my term) has its own journals and funding sources, its own jargon, conferences, associations, and communication links not available to outsiders.

6. The Need for Individuation. Again this relates to funding and career needs that promotes short term egotistical behavior. In order to have a career in academia graduate students need to be different (appear creative) in some way from their classmates in order to get awards and recognition. As an assistant professor one again needs to be different and claim some specialty for one's own in order to get tenure. As a tenured professor running a research lab one again needs to be different in oder to keep the money coming in. None of this promotes collaboration or long term research into a common area. Lenat then points out that all the easy ground breaking work that could be done by an individual in a few years has already been done. The work remaining is harder and longer term.

5. Automated Evolution. Another time waster is the attempt to simulate evolution in the hope that intelligence will emerge. The problem is that what emerges is always something simple that has been essentially designed into the simulation. This is now the common simulation in alife but again I must agree that this approach will not lead to anything significant especially if they use artificial neural networks or small bits of algorithmic code. These simple solutions lead to publishable papers but they do not scale up into anything really worthwhile.

4. More Overscaling-Down: Automated Learning. This is the problem in which some sort of A.I. or automated learning appears more sophisticated than it really is. All programs have a built in structure that is not apparent in a publication. Programs can be biased to achieve certain desired solutions, essentially cheating on the problem by making a special case easy for it to solve in order for the author to get published. So many models lack generality and thus do not provide general solutions.

3. General Language Understanding. This is similar to the above AI impediment except applied to Natural Language Understanding instead of machine learning. To quote Doug Lenat: "NL researchers -- have watered down the problem to the point where they 'succeed' on every project."

2. We Can Live Without It. This AI impediment relates to the problem that people do not really need good AI in their lives. Without a compelling need funding and talented researchers will be scarce. Lenat mentions computer gaming as not needing good AI because most computer gamers are content with the simple behaviors of non-player characters as they now exist. Being a computer gamer myself I have to agree but only up to a point. Most online gamers complain about the "grind" in MMORPGs and after the novelty wears off and leave the game and after a while leave gaming altogether. So I think their is much potential there if online gaming companies want to grow by keeping their customers.

In regards to the Real Simulated Intelligence (RSI) approach shown on this site another real need is to understand how the brain works in order to better understand how various brain illnesses work. Where is their locus? What form of neuromodulation is affected? Is there any strategy that can get around this brain defect?

1. There is one Piece Missing. The final impediment and number one on Doug Lenat's list is the need to get researchers working on the hard long term problems like he and his team at Cyc are doing. Implied here is the thought that serious researchers need to get out of academia because no one in academia is doing this needed work for all the reasons outlined above.

Edit (July 5, 2008): I think the reason so many "time wasters" exist is due to the sheer desperation people have in not having any clue on how real system wide intelligence works. Consequently, they will try anything no matter how far fetched.