Stanford AI Class Lecture Notes • Hacker News Board

Re: Stanford AI Class Lecture Notes

#21

Post by Mvandenbergh » Mon, Oct 10, 2011, 11:39 AM UTC

guscost wrote: post #15

Couldn't this understanding of intelligence limit the ways in which we can describe and emulate the activity of intelligent creatures? For example, the interfaces and processors are all very clearly defined and separated in those diagrams. Unfortunately, natural intelligence does not seem to work in the same way. The inputs to a real human do not get processed in the same places, even when they might be coming from t…

That's true, but most of modern AI is about designing useful intelligent agents, rather than creating artificial consciousness.

Re: Stanford AI Class Lecture Notes

#22

Post by Mvandenbergh » Mon, Oct 10, 2011, 11:43 AM UTC

dirtyaura wrote: post #14

xarien wrote: post #4

Google still has a very long way doing an adequate translation between a language such as Chinese and English. I'd be surprised if Google was taking such a pedestrian approach in solving that particular problem. As it stands, languages such as Chinese are intrinsically implicit in nature. In fact, the more adept at the language, the more you can express with less. If you follow the literature back a couple thousand y…

Statistical translation that Google uses works quite well to translate between most Indo-European languages. However, it has problems in other languages like China and Finnish, but they are not necessarily insurmountable if you combain statistical approach with some domain logic and a lot of material. Developers of Google Translate described to me that Finnish is causing problems because the excessive inflection [1]…

One of the reasons that the EU documents work so well is that they're painstakingly written simultaneously in all the official languages. There is not a master copy from which other translations are made. This also means that if, say, the German working group writes a clause that is only approximately translateable to French, both groups will have to re-write their versions so that there is as close to possible as an exact correspondence between them.

Other possible corpuses are mostly literary, which are of course subject to significant rewriting for stylistic reasons.

Re: Stanford AI Class Lecture Notes

#23

Post by guscost » Mon, Oct 10, 2011, 3:16 PM UTC

Mvandenbergh wrote: post #21

guscost wrote: post #15

Couldn't this understanding of intelligence limit the ways in which we can describe and emulate the activity of intelligent creatures? For example, the interfaces and processors are all very clearly defined and separated in those diagrams. Unfortunately, natural intelligence does not seem to work in the same way. The inputs to a real human do not get processed in the same places, even when they might be coming from t…

That's true, but most of modern AI is about designing useful intelligent agents, rather than creating artificial consciousness.

Well, that's a cop-out.

Re: Stanford AI Class Lecture Notes

#24

Post by aapl » Mon, Oct 10, 2011, 4:12 PM UTC

jcarden wrote: post #8

LOL, here's what I just got when I tried to log in 'Ooops Our servers are off having a quick coffee break. Wait a second and refresh the page. If you still get this message, we apologize and ask that you try again a little later.'

If you repeatedly get this message, remember to click reload. It seems that it can get cached in your browser.

EDIT: It seems that it isn't just that and the site is just flaky (overloaded, I guess).

Re: Stanford AI Class Lecture Notes

#25

Post by Udo » Mon, Oct 10, 2011, 4:46 PM UTC

guscost wrote: post #15

Couldn't this understanding of intelligence limit the ways in which we can describe and emulate the activity of intelligent creatures? For example, the interfaces and processors are all very clearly defined and separated in those diagrams. Unfortunately, natural intelligence does not seem to work in the same way. The inputs to a real human do not get processed in the same places, even when they might be coming from t…

Brains are a problem that is drastically out of scope as far as AI is concerned. You're absolutely right, these are just very simplistic models and examples, meant to convey the general AI jargon. This isn't intended to apply to "complex" systems; for example the idea that you have a very basic sensor, a black box, and a very basic actuator works well for a specific problem set, but it's not intended to work as a model for brains (or even parts of brains).

I believe the source of your disappointment is a matter of overall expectation about what AI research is intended for. The objective of AI is not to create intelligent beings, it's to model and create programs that solve narrowly predefined problems. AI as a field is not at all identical to AGI (=artificial general intelligence). Whenever you're talking about brains or things like "common sense", that's AGI. Over the years AI research has produced many good models for single components of our mental subsystems though. But researchers have not actually concerned themselves with AGI until very recently.

Re: Stanford AI Class Lecture Notes

#26

Post by xarien » Mon, Oct 10, 2011, 11:57 PM UTC

dirtyaura wrote: post #14

xarien wrote: post #4

Google still has a very long way doing an adequate translation between a language such as Chinese and English. I'd be surprised if Google was taking such a pedestrian approach in solving that particular problem. As it stands, languages such as Chinese are intrinsically implicit in nature. In fact, the more adept at the language, the more you can express with less. If you follow the literature back a couple thousand y…

Statistical translation that Google uses works quite well to translate between most Indo-European languages. However, it has problems in other languages like China and Finnish, but they are not necessarily insurmountable if you combain statistical approach with some domain logic and a lot of material. Developers of Google Translate described to me that Finnish is causing problems because the excessive inflection [1]…

I agree to a point. For one, I would not group Finnish and Chinese together. Now, I can't speak much for Finnish, but I am positive a statistical approach to Chinese provides very poor results. However, with that said, I do agree that domain logic and a large relational data set is a good approach. Essentially relationships between words and symbols need to be established with false positive and false negative associations broken down and re-established in near real time to come close to a perfect solution. Basically, the problem is too expensive to solve.