Neural Nets

Post Reply New Topic

I'm working on a computer simulation of virtual life- I know it's been done before but I have a few ideas I think are new. Anyways as part of the research for this project I've been reading a lot of papers on neural nets, and honestly I'm starting to think that "real" AI is not as far away as people tend to assume. A lot of the research before this decade wasn't that interesting, but in particular David Vogel from the Dominican Republic has a lot of really good ideas and although I haven't finished digesting his work I can already see how a really powerful neural net might be built. Of course there are huge practical considerations of teaching the AI, figuring out how to motivate it, giving it sufficient sensory input to not be a vegetable, but as I was thinking about it today I started getting really worried about where this line of research might lead. I mean, realllly worried. In fact I think that if I had access to a powerful enough computer I could possibly create a really intelligent AI in a matter of a few years of work.
But maybe this is not something I should contribute to. Not that I think there's going to be some kind of hollywood style Terminator 2 situation, but.. the first checkers program I built, could play checkers much better than myself, and that didn't bother me since a checkers program is just a machine which plays checkers, and now checkers is a solved game anyhow. But if I contributed to the development of a machine which was way smarter than any living human...? Once you get the basics down scaling up only requires computing power, which will eventually become available. I don't need to go on, you are all smart enough to see where this line of thought goes.
I don't know, it makes me nervous. Not nervous enough to stop the little project I'm working on now, which will not feature very large neural nets, but it makes me nervous. The roomates where I live watch Rock of Love every day and they have no concept of what goes on upstairs and what it's going to mean someday.
It's stupid, right? This whole post probably belongs in the realm of science fiction. But then again...

Its unavoidable that we will trigger a machine evolution that will replace us in enough time.

why do people automatically assume that a true AI becomes sinister, or a threat to humans?

It want be sinister. It will help humans. But in time machines will evolve and expand over the universe in a way that biological systems will not beable to. Thats not a threat to humans because human largely want beable to fill that niche in the environment.

Oh, silly me. Here I was, thinking it's just an irrational fear of the unknown

While this maybe isn't 100% spot-on, it's an interesting read: http://www.gamedev.net/reference/articl ... cle771.asp

I guess it's pointless to worry about it at this point.

Because the moment a true AI is exposed to television, it will either go totally insane or be convinced that humans are psychotic, schizophrenic sociopaths who must be put out of their misery. Or both. That is, after all, the only logical conclusion any machine intelligence can come to after observing Superbowl Sunday - and in particularly the adverts.

It seems Kubrick and Clarke got it right with HAL.
Any intelligence designed by humans will inevitably acquire our flaws.

I really hope this isn't true

I am employed at an AI research centre and I cannot understand why people always think AI would mean to replicate human behaviour using machines. We already have got humans, we do not need to build or research machines to replace them (at least not entirely).

All we do is researching to make computers being able to do things like

image recognition, computer vision
natural language processing
machine learning, deriving patterns from a small set to apply it to a larger set of problems
...
and more like that that is just more than preprogrammed or scripted actions.

Building computers that want to live, talk, eat, fight, have sex, care for their survival and for each other and compete with humans is absolutely pointless. Why should anybody want to do that? Give me a reason, please!

There are some characteristics in humans interesting to have in computers but any of those characteristics we also find in most animals. Bees or ants for example live in social communities, they practive teamwork, they see, they navigate, they walk using legs (bees even fly), they react to sensor stimuli and they adapt genetically to changing environments by evolution.

A machine smarter than humans? What is smart about a human? Reacting most appropriately to a given situation? What would make a machine facing human like situations? Getting along with life? With the job? With other people or machines? What is smart anyway? What is human intelligence? It is all about comparing humans with other humans and we measure intelligence using bell curve statistics. Comparing computers or other built machines with humans is not like comparing apples with pears, it is comparing stones with pears.

That's just it, there are AI nut jobs who are obsessed with emulating humans and animals. So you can't exactly blame the public for thinking that way. They also use scare tactics (as if you would design something that wouldn't ever run out of juice or you couldn’t turn off) at the same time as expecting people to give them cash for their crazy ideas. Quite bizarre.

I think there are quite a lot of delusional people involved. They want to be the one, who’s cracked it. But there isn't it. It is too big to quantify.

the_falling_frog what make you think you know how much time it is going to take? I wouldn’t focus on that that is a distraction.

A lot of the famous projects seem to involve quite large areas. What is this with trying to do everything at once? It makes more sense to me to work on a component of intelligence that could be useful. Such as judgement theory through things like computer go. That would be a useful application in order to make systems more efficient without needing greater 'capacity'. Capacity is joke if it is not used efficiently. Humans don't have a huge capacity, but they make better decisions than a computer can.

Last edited by 0_equals_true on 29 Aug 2007, 7:21 pm, edited 1 time in total.

I guess the variable you people are not considering is self replication. I have a small CNC milling machine that I could theoretical copy most its own parts with. Which basically means the day will come when it will be trivial for such a machine to completely copy itself. And when that happens you will see machine evolution.

http://www.wrongplanet.net/modules.php? ... ht=destroy

What makes people think they are God's gift to the universe? Get over yourself people a machine will be digging your bodies up to power its fuel cells in the future.

Machines have no reason to replicate themselves. They have no need to survive, they do not compete for scarce resources, they have no intentions or objectives. Computer viruses already can copy and spread themselves, so we already have got self replication.

There are concepts like genetical programming where algorithms or functions compete for a given scenario, there is intentionally triggered mutation from one generation to another and the candidate with the lowest error value wins the current iteration. So we already have got the concept of evolution in AI.

There already are complex multilayer artificial neuronal networks and they are able to learn complex functions (i.e. by gradient descent learning), for example driving a car by watching the road.

Although all of those concepts sound rather interesting and as if they would mean human like AI is just around corner: This is absolutely untrue. Even the most simple animal brains outperform any possible hard- or software solution at a ridiculously extreme magnitude.

For learning about evolution and its possible effect to AI I can recommend the book

"The Meme Machine by Susan Blackmore" and

"The Selfish Gene by Richard Dawkins" is also quite interesting as a background knowledge.

Self-replication is the only objective for the machines and its the only objective for people. All it takes is for it to start and it will start.

Yes, I agree with you on that. The issue is that we, as humans, have to build such machines in the first place, before they will start replicating themselves. They will not suddenly start doing that because of better AI.

Last edited by byrlawson on 30 Aug 2007, 10:59 am, edited 1 time in total.

Where do you work at? I'm of course not going to create any form of self replicating machine, but in the course of my research for this project I'm making some interesting connections between pattern recognition and recursive data compression. I'm going to construct a heirarchical network of small pattern matching modules in which each one looks for matches with a certain number of patterns (or maybe just one pattern per unit? whatever works better) and by making these modules send signals when they detect something familiar perhaps the network can derive deeper relationships between the simple obvious patterns of input. I think the first thing I'm going test this on is some form of artifical visual cortex- from the looks of it this kind of processing doesn't require much learning but is mostly hardwired in the brain, well from what I've found on google anyways, so this might be a good place to start. I'm still trying to decide how training will take place, but this type of network might require feedback and not work well with simple backpropagation schemes. The best way to train it may be some kind of reward/punishment system where I try to teach it to recognize simple shapes and so on... and when it gives the correct answer the recently fired connections get stronger. Which of course involves some sort of working memory. I'm still sorting it out. I think some of the structure of the network I'm just going invent from imagination based on what I want it to do, only the deeper layers will be trained. I realize that I'm being vague but is this a line of research that others have tried already? It would actually surprise me a little bit if some of this stuff hasn't been tried yet- maybe it has and I just am too out of touch to have heard about it.