wow! Projects in AI

One of my lecturers had that happen to him, apparently. A GA evolved a program which he cannot understand.

It actually happens a lot with complex neural networks. It's not easy to understand why they work. Or don't. And you usually apply a neural net to a problem you don't allways have a direct solution (otherwise what would be the reason?).

But unfortunately I don't have any good stories about that.

[edited by - Cesar on April 17, 2004 7:22:07 AM]

I read a similar story in the famous books of John Koza (Genetic Programming) where he designed circuits for high/low pass filters that worked better than any other in existance and nobody knew why. He did a lot with analoug circuits.

Personally, I am working on a Genetic Programming system where everything is executed form a Lua script. So this system can be VERY cusomizable. I am hoping to do some cool stuff with the stock market when I am done. And also agents in a fighting type system or some other problem that is difficult to express directly.

Dwiel

My home page!!!
Find out about my diy LCD projector and programming projects!

I read an article in a recent edition of New Scientist that highlighted work funded by NASA into intelligent microphones. Traditionally, in environments like space, underwater and the military, throat-mics were used to directly detect vibrations in the air in the throat and in the facial and throat tissues that were produced by speaking. However, while this is an improvement over using a traditional mic in front of the mouth, it is still unsatisfactory, particularly in space where the inside pressure os a space suit is quite low, so as to prevent it from becoming very stiff in a near vacuum.

So, a research team is working on a system involved electrodes in contact with the throat that directly detect the nerve signals sent to the vocal muscles. These are classified (against a training set for that speaker) such that the system can understand the words being spoken. So far, it can learn about a dozen words (like STOP, GO, LEFT, etc). As it happens though, the person doesn''t actually need to speak. It''s sufficient to just mouth the words and in some cases, it can even detect just the thought of saying the word. If this could be further developed, one could have a whole conversation with someone without ever moving their lips or emitting a sound.


In a similar vein, I found a research article on Friday (related to some of my research) that was a report of the use of a bank of artificial neural networks to detect the component of EEG signal generated when a person thought about moving a finger. This ''filter'' of ANNs could detect the signal in the midst of usual EEG signal (which is extremely complex and filled with dynamic noise). Very impressive and very important when you think about brain-mechane interfaces. Prosthetics would become useful and would require the retraining of muscles at the end of the severed limb. Humans could control machines through thought. Very sci-fi, but very real too!

I hope this generate the ''wow'' factor you are looking for!

Cheers,

Timkin

You successfully wowed me!

one question I have about the NASA microphone thing is how it can do it without you moving your lips... aren''t some noises determined by where your lips are? Such sounds as ''eh'' as in elephant and m. try making the ''eh'' sound with your lips closed and you get m... How would it distinguish between these if you never moved your lips? Maybe they were using the brain electrodes/wave detection for this...

Very interesting. I''ve also heard about they have a monkey with electrodes embedded in his brain and he controls virtual things with it... like pictures on a screen. They are working on getting him to be able to use it to use another arm. Very cool stuff.

Dwiel

quote:

---

Original post by Tazzel3D
one question I have about the NASA microphone thing is how it can do it without you moving your lips... aren''t some noises determined by where your lips are?

---

The noises we make when speaking are produced by complex movements of many muscles. The NASA system I described doesn''t work by detecting air vibrations or muscle positions, but rather the electrical messages sent to these muscles by the brain. It''s not necessary for the muscles to actually move to produce the speach, since the speach must be encoded in these electrical signals. Just think of a telephone line. The words you speak are encoded electrically, but you need to know how to decode it to understand the speach. The NASA system doesn''t actually decode the muscle signals, it simply matches signals to words, based on a training phase for that particular user.

These signals are obviously not the same signals as those that cause the muscles to contract (or perhaps they are, but just not strong enough to cause the contraction), but they still exist, even when the muscles don''t contract. The electrodes are designed to sense these signals and pass them on to the filter/classifier, which determines what word is being spoken.

I hope that makes it more clear.

Timkin

Yeah, that makes more sence.

Thanks

Dwiel

My home page!!!
Find out about my diy LCD projector and programming projects!

I was working on a PhD which involved using neural networks to model the early visual processing system. Some of the emergent properties of the network - that is, behaviour which was not explicitly programmed into the neural network - demonstrated behaviour which was very similar to visual artefacts in the human visual pathway e.g. mach-band effects. Meaning that the artificial neural network was subject to the same optical illusions as the human visual system. That was very encouraging.

@ Timkin:
Yeah I have read the same article, apparently your mind sends signals to your throat muscles whenever you think about words. All you have to do is think about the word, and then your mind sends signals to your throat incase you are going to speak them or something. So the mike can pick up on these signals and translate them to speech without you talking! Kind of a hack mind reading type of thing. The problem lies in not detecting the signals but accurately determining what the signals are trying to generate for words.

I used to work, not too long ago, with a guy that used to work at IBM. His name was Dave Modlin, and he was like my Mentor and idol all rolled into one. Anyways, he told me this story once, similar to the one you mentioned with the circuit boards.

Basically, it was an AI that was designed to produce circuit boards to do various things. It began simply enough, and then began to make circuit designs that, though feasable, would only be possible with manipulations beyond three dimensional space (like having two different bits of circuitry exist at the same point in space was how he explained it, but it was a simplification), and reused other parts of the circuitry in ways that was never before imagined (in essence, feeding back into the system to produce alternative results).

I''m not sure if it was just a fish tale that one Senior programmer told to his newbie junior programmer to "wow" him or play on his nievity, but either way I''ve always remembered the story.

How about the work done in evolving layouts for Freely Programmable Gate Arrays (FPGAs) at Sussex Uni. One particular network produced (and we''re talking hardware here, not software) the required output, but oddly enough had a section that was not connected to the rest of the array (in other words, the evolutionary process had formed two arrays...one large and one small). When this smaller section was removed (because it was thought to be an unecessary set of gates), the FPGA failed to produce the required output. This puzzled the researchers involved, until they realised that the two sub-networks were communicating via the local electromagnetic field generated by the activity of the gates. Talk about ''freak me out''!!!

This same effect has been recently reported (about 4 weeks ago) as having been found in neural tissue from mice. Information can be passed across a gap between tissue that is mechanically separate, via the local chemical and ionic fields, as opposed to only neuronal connections (synapses).

Timkin

I think the things that wow people, are the things that just about everyone knows how to do, yet the computer can do it better. Of course anything related to games is a plus too.

I''ve used my Tetris Agent at a few events where we had to demo something for school... it always drew a lot of attention both from the students and from the adults.

TD-Gammon I think would impress a lot of people too. Something like that is the way to go IMO. It also makes sense to have more AI components (like the example above of a system grading itself) and anything with anthropomorphic qualities.

This is in response to the first article mentioned, where the researcher was evolving circutry.

Timkin: You neglected to mention this (I read about this guy as well):

Some of the solutions were also taking advantage of impurities in the materials that phyiscally made up the chip.

Very spooky.

Will

Some of these posts are pretty cool! Does anyone have any links to articles or info about AI generating code and/or circuits that work amazingly well but confuses the hell out of the scientists/engineers trying to figure out why it works?

Snowman | Ardaxus | Nathan
---------------------------------
Tangle - Persistent World Toolkit
tangle.netlab.com.au

I dont have any links on me, but I do know that John Koza, the one who ''started'' GP, has a book out which talks for a couple hundred pages about doing just this. I think it is volume 3 of his books. Not sure if you have access to it or are motivated enough to get a hold of it, but I thought I''d let you know.

Dwiel

My home page!!!
Find out about my diy LCD projector and programming projects!

Scientific American Magazine: Report on Evolving AI - PDF 6MB

Read it, it talks about AI that evolves circuitry, and some of its current and future applications and predictions.

Truly a glimpse of the future...

---

[Hugo Ferreira][Positronic Dreams][Colibri 3D Engine][Entropy HL2 MOD][My DevDiary]
[Yann L.][Enginuity] [Penny Arcade] [MSDN][VS RoadMap][Humus][BSPs][UGP][NeHe]
"our stupidity allways comes back to bite us in the ass... in a white-shark sort of way..." - Prozak

In the nineties, researchers were working on a very impressive AI. It had the potential to automate the work of the Pentagon, and would be a much more reliable and efficient manager of the nuclear weaponry than the President. In case of an attack the AI would calculate the most efficient response and execute it. It was also going to network all the branches of the military to better coordinate responses. I believe it was supposed to go online in 1997, but something must have happened - I haven''t heard anything more about it since then.

blah! poor attempt at humour... troll!
August 29th, 1997, Skynet, Terminator 2 - Judgement Day...

---

[Hugo Ferreira][Positronic Dreams][Colibri 3D Engine][Entropy HL2 MOD][My DevDiary]
[Yann L.][Enginuity] [Penny Arcade] [MSDN][VS RoadMap][Humus][BSPs][UGP][NeHe]
"our stupidity allways comes back to bite us in the ass... in a white-shark sort of way..." - Prozak

quote:

---

Original post by Prozak
blah! poor attempt at humour... troll!
August 29th, 1997, Skynet, Terminator 2 - Judgement Day...

---

I know... I posted it at 10 or so in the morning after being up all night. Didn't mean to be trolling though (notice the grinning face in the header) :/

Anyway... The other examples were great and a bit spooky sometimes. Anyone have more or has the well run dry?

[edited by - frostburn on April 22, 2004 2:09:23 AM]