[DavidOlmsted]

I am not yet exactly sure what intelligence is beyond being something the brain does but scale is certainly one component. Intelligence does depend on accessing a huge amount of information. The question then becomes whether a sequential algorithmic architecture can access all the needed information in anything like real time or whether an inherently parallel architecture will be required?

I suspect the limits of sequential processing have just about been reached due to speed of light limitations in a CPU that does not melt. This means the whole sequential A.I architecture approach is a dead end. Any radical new technology that shrinks information processing further will go into the quantum level in which case the uncertainty inherent in that level means the end of finite state machine (finite automata) usage anyway because discrete states can no longer be defined.

Instead truly intelligent systems will have to be built within an inherently parallel architecture and that means brain-like including memory prediction. A general purpose parallel processing computer has not yet been successful developed despite much research. Parallel processing only works when the problem being simulated is inherently parallel to begin with.

[csaba_trucza]

Hi David,

I almost agree on your assertion that "the whole sequential A.I architecture approach is a dead end".

However I don't think it's because of the limited processing power. More improtant - I think - is the central role of time in the way the brain works (as I understood from the On Intelligence book).

I was just thinking how could one detect temporal patterns (using some artifficial neural networks).

One approach is to use a sequential architecture and simulate the passing of the time. This probably means an artifficial granularity of time, which seems pretty unnatural.

One could first go to quasi-paralel architectures (like threads or such), but this will quickly bring down most computers (imagine running 1 million threads in parallel. Few computers can do it). And only a real-time system would permit an undistorted realization of the passing time.

True parallel architectures which will fit more naturally to the problem as you don't need artifficial engineerings to factor the passing time in the model.

[Joshf]

I agree with you csaba_trucza that this is the reason why the sequential AI architecture will not stand the test of time. The only thing we can do now is simulate a time-based system by breaking simulating a step by step process in the entire network. It is a shame that this must be done. This has always been known however. We should have seen this since they knew that neurons have a downtime between firings. I just hope that it is not too much of a hinderance to the overall goal of AI.
_________________
Josh Ferguson
Undergraduate, School of Mathematics
Georgia Institute of Technology