OSDev.org

bubach wrote:Thats basically what I believe too, but thats just because the "internal model" as yoda calls it is basically hardcoded information about for example corn and fields. No matter how much code you pour into that to make it be the perfect corn picking robot it's never going to become self aware or able to preform well on other general tasks. Thats how robotics and AI work right now, and I think it's a waste of time and resources. We'll never be able to program for every possible condition and even with self learning additions to the base code, it feels mostly like ugly hacks to me.

No, that's not it at all. They are based on data structures which change their structure and contents (hence, the modelling is adaptive); it's not code-based at all. I've mentioned things like genetic algorithms and ANNs several times here but, apparently, no one seems to have bothered looking them up.

Love4Boobies wrote:No, that's not it at all. They are based on data structures which change their structure and contents (hence, the modelling is adaptive); it's not code-based at all. I've mentioned things like genetic algorithms and ANNs several times here but, apparently, no one seems to have bothered looking them up.

Yes but the somewhat useful ones out there are mostly still geared towards a very specific use case with some stuff coded in.

I have yet to see a (successful!) fully adaptive robot that tries to learn everything from the ground up with no knowledge at all programmed in from the start. If there was, that AI would be "finished". I know of some experiments like this, but they spend like 5 years making those self learning robots make out the difference between a circle and a rectangle, or making out cats in photos - while humans have learned to speak, walk and lots of stuff in the same time frame.

So while there is research in the right direction, my point is that most useful robots that preform a specific tasks today are "cutting corners" because research haven't gotten to were it needs to be.

That's not entirely true. For instance, I wrote an optical character recognition (OCR) program using ANNs. It had nothing coded in and I trained it from the ground up (there are several ways to do that---either supervised, like humans, or not). The two biggest problem with ANNs are (a) that they take huge amounts of time to solve complex tasks due to most hardware, which isn't parallel enough, and (b) that they take huge amounts of memory. They are thus incredibly expensive for complex tasks. We're also learning a lot about ways to optimize the networks while training in such a way that it won't affect future learning too much (although I believe this came from graph theory, recent research shows that it's extremely like that our neurons die so that we become smarter because our neural networks can take certain shapes).

Love4Boobies wrote:Exactly, you're not being patient. Think of Fermat's last theorem---it took the world's greatest mathematicians 358 years to solve...

358 years of patience – I won't live so long .
Look, I don't state that AI is impossible at all. I'm not Penrose. Yes, this question is arguable. The only thing that I want to explain is that AI task is frequently oversimplified. It can't be compared with so particular mathematical problems as Fermat's theorem or N vs NP problem. These problems may or may not be solved in any moment of time, depending on depth of penetration of mind of specific genius to the task. But AI is not theorem. Potentially I see ways to approach to the task, but in practice I see that it will require much more time than it is estimated by many AI researchers.

Love4Boobies wrote:So given that the AI problem is one that we don't fully understand yet, how can one hope for a solution in just a short amount of time? We will probably not live to see it but that doesn't mean it won't happen.

That's what I'm talking about!

Love4Boobies wrote:I've mentioned things like genetic algorithms and ANNs several times here but, apparently, no one seems to have bothered looking them up.

I spent much time with adaptive neural nets but completely disappointed with this concept. NNs are not able to solve any task that could not be solved more efficiently with common algorithms.

Love4Boobies wrote:For instance, I wrote an optical character recognition (OCR) program using ANNs...

I talked many times with authors of FineReader OCR (we studied together at the University). They used NNs in early versions of FineReader for OCR task but soon rejected NNs because of their inefficiency and unreliability. That's commercial project of high quality!

Fair enough, I thought you hinted it was impossible.

Yoda wrote:These problems may or may not be solved in any moment of time, depending on depth of penetration of mind of specific genius to the task.

It may or may not be the work of a genius. I like Knuth's metaphor about the evolution of computer science: It's like building a wall, where many people contribute a brick each; while one single brick may not be a huge deal, after 5 years the wall is significantly taller. We've mentioned Fermat's last theorem; that was proven only because many other problems were tackled (the last of which were the epsilon, Serre, and Taniyama–Shimura conjectures).