Site hosted by Build your free website today!

Artificial Intelligence

Back Home Next

Yes now I am adding this chapter!


I think people may mistake "artificial intelligence" for something that is a threat and superior but because it is artificial it lacks "humanity" so that there is a nice safe Achilles heel to it. The reality is different.

The question of what intelligence precedes this. We have measures and doubts on their relevance. For example IQ tests are often considered to be culturally biased. Also, our metrics on parameters is incomplete - e.g. a dancer may have "mobility intelligence" but solving a rubix cube won't show this capability.

Humanity as a whole has progressed chaotically. We have wars, dictators, skirmishes, massive human death - so if we were an intelligent species how did it take us so long to invent a light bulb?

OK I'm being facetious but what is "intelligence?" - how can we know if we have created it in silicon?

My guess is that we cannot know as the term is without meaning. We can only look at functions. A small part of the brain, for example, can give rise to musical genius and yet appear structurally to be the same as any other person. Some people have massive memory capacity but cannot perform simple social tasks. Intelligence is therefore a weak term and it is better to describe functions that can be measured.

Neural Networks are an interesting parallel to how we think a brain structure may work. However it only acts as a small sub component of the hypothetical musicians mind. The question I pose is what are the building blocks at a functional level.

Analogy - transistors make up a CPU but it has logic elements like AND, NAND, OR, NOR, XOR etc as building blocks. How can we design a Neural Net without this foundation?

The net is a simple structure with good memory performance but unable to solve complex tasks like multiplication. The Romans had a number system that foiled them on the same task - so are people Intelligent or is it more a question of luck, time and adaptation?

This further suggests to me that more fundamental building blocks are needed. After all, multiplication, in this analogy, is a procedure and requires a base system. Well with integers;

        x y = exp[ loge {x} + loge {y} ]

        x / y = exp[ loge {x} - loge {y} ]

These equations work with all numbers. Also replace "+" with "" and "-" with "/" and you get new binary operators - these lead to new numbers also.

So how can a Neural Net (NN) accomplish such tasks - well it can't. A NN can only be trained to see patterns. They are excellent at this but cannot develop new mathematics. Or can they?

Multiplication in any system is a procedure. In binary there is a carry sign. Procedures are sequential and can be programmed. However this approach is at odds with these considerations of what "intelligence" is, if it exists, or how to define it for replication by a silicon machine.

Maybe working the procedure out is "intelligence. Maybe applying a procedure is just mechanization. I think the Greeks invented decimal 10 as a number system - how could anyone multiply with such clumsy number representations as I, II, III, IV, V, VI, VII, VIII, IX, X, etc. What a stupid system this was! And yet Romans maintained it. So can we say humanity has intelligence based on such a dumb number system or are the objectives of those that subscribe to that old number system at issue here? Maybe some people benefited by having "expert capability" in the use of a poorly founded number system.

So we now also need to consider "deception" in intelligence. It may well be directly related as a survival mechanism - after all camouflage is a form of visual deception. Politics is a verbal one.

Our first problem is not having a logical definition for intelligence so if we witness it how will we know it's there? SETI has this same difficulty - what signals would be sent - a binary on/off sequence is also dumb. At least send a prime sequence - this proves a knowledge of numbers.

So this sets the background for NN design. We can only measure function and still lack key elements of knowledge here. We can make simple NNs on a Field Programmable Gate Array (FPGA) to make use of parallel processing (PC CPU's and DSPs are useless for this except to debug a code list) but we soon reach a limit in capability. This is problematic - the analogy again is if you just have a hot heap of transistors in your hand - without a circuit structure they can't be used. We lack the building blocks!

Neural Network Structure
Back Propagation
Competative Training
Consensus Training
Multiple NN Development


Return to Ian Scotts Technology Pages

Ian R Scott 2007 - 2008