Quote:
Originally Posted by Buffy
Great topic Twit! One of my faves.
I will tell you that I don't think we have a good enough handle on the definition of the words wiki is using to define consciousness, and as a result, no, we can't define "consciousness." Can you define sentience? How do you demonstrate "self-awareness?" All of them are so subjective that they come under the rubric of Justice Stewart's "I can't define it but I know it when I see it."
|
It's an interesting exercise to get people to define what "learning" is, which seems on the surface a much easier definition. They quickly run into problems either leaving things out that we'd call learning, or by including things (such as a car engine "running in") which we wouldn't call learning.
Quote:
|
We're a lot further along now technologically than we were when Alan Turing broached the subject, but that's not saying much. He basically seemed to say: the only way you can tell is if you sit down in front of it and a real intelligence and can tell no difference.
|
It's always nice to see a correct description of the Turing Test
It's a personal bugbear of mine that many people seem to think that you only talk to a single agent, and the question is whether it's a human or machine. This leaves out the game-playing side to the test where the human will be encouraged to use the full scope of their intelligence if they think the machine is likely to "beat" them.
Quote:
As a practical matter, my personal belief is that the implementation "intelligence" will be far simpler than we think, but that creating it is an evolutionary (as in letting neural networks build themselves) rather than algorithmic.
Fuzzily,
Buffy
|
Oh dear. Again I have chosen my moniker appropriately. While I have a long history in AI, I'm, shall we say, less than impressed by Neural Networks. As a learning technique they are, I believe rather weak. The problems they solve are limited to simple problems that can be expressed in a zero-order logic (*). And they are not able to make use of background knowledge, meta-knowledge instructing them how to learn. They try to learn everything from a near true "scratch", which I do not believe humans do. Modifications to neural networks to make them learn in these ways, I believe, changes them so much that they are no longer neural networks.
I think that neural networks are actually a dead end in learning research, and that they will eventually die out. They are popular. And that's not only because people think it's "cool" to simulate the human brain or suchlike. But also because they do work quite well on simple problems and there are a lot of simple problems around. But, while the current huge amount of effort goes into NNs, less work is being done on the research to invent and develop techniques that will solve the more difficult, more "intelligent" problems.
However, even for the simpler problems that NNs do well, I still prefer what are called Support Vector Machines for learning. You mention NNs "evolving" themselves to find better solutions to problems. I'm not sure if you are referring to neuro-genetic systems here, but these are systems that use evolutionary computing techniques to optimise their weights to solve problems. Or, in simple language, they sort of search around for a good solution to a problem. With support vector machines, the learning problem is transformed into a form that can be solved optimally using mathematical optimisation techniques. That doesn't mean that they always do better than neural networks on every problem as something may be lost in the transformation. But, if someone put a gun to my head and forced me to say whether I thought NNs or SVMs would work better on an undescribed problem, I'd go for the SVM any day.
If anyone looks at the book "Machine Learning: An Artificial Intelligence" approach, vol 1 (Michalski, Carbonell, Mitchell), and look at the kind of problems people were trying to solve over 20 years ago, and look at the kind of problems people are trying to solve with neural networks nowdays, I suspect my point will be made.
(*) Note that it's possible to convert
simple problems typically expressed in higher order logics into zero-order logics such as the fixed length vectors of numbers typically passed to neural networks as input. But the methods for doing this completely break down when more complicated problems are attempted. Complicated problems that humans solve.
