Robotics and AI

So how close are we to building robots like those in the Asimov scence fiction books? I hope you all know the ones I mean. Humaniforn, walking, talking, thinking robots. I suspect a long way a way. I've worked in the computer industry (computer manufacturing), for the past 30 years, mainly in the software development and design arena but I have never worked with AI (Artificial Intelligence), but would love to.

Does anyone have any experience with AI that they could share with us.

I never quite understood the term AI. Intelligence is intelligence, how can it ever be artificial. I'd like to see the term changed to MI for Machine Intelligence.

The main issue for AI seems to be visual recognition and interpretation. Until a device can view something and recognize it easily then everything else is largely unimportant. The real issue is interpretation. Think about what happens when you look at an object. There is 3D, colour, lighting, position, and what if something is partially hidden. You know that if you see someone you know from the angle of part of the back of their head, then there is a good chance you will know who it is and fill in the details. Telling a computer to do the same thing is enormously complex. The methods for collecting visual information is easy with machines as it is with our own eyes, but it is the brain and memory that really does all the work.

I was in a meeting about 2 months ago with Jim Gray, Head of Microsoft Research here in Silicon Valley and I asked him his opinion on AI. He said that little progres will be made until we can make a megamemory (can't remember the axact term now). But essentially it requires that we collect all the useful data in the world and put it in one place. We then need a computing engine that will be fast enough to process the data within reasonable times. The idea here is that you can ask it any question and it will respond rapidly with an appropriate answer. A bit like a superbrain. I suspect before we reach that point we will have subsets.

Well that's a start. I'll add more ideas as I think of them. Please add your own thoughts and ideas.

 

Log in or Sign up to hide all adverts.

Silicon Valley! Hello, Homey! This is my native turf. I was born here, I plan to die here.

When I was with the MGM project, they introduced us to Eliza. As an early experiment in AI, it was programmed to pick out key words and phrases and then select from a list of appropriate responses. It was good, but easily confused.

As far as the phrase itself, there can be some problem with the definition. Normally, it's defined as being when a human carries on a remote conversation or some such activity with a machine, and has no idea that he is dealing with one. With a list of appropriately vague responses, however, I can set up an answering machine to make a caller believe that I am actually home for awhile, provided they don't ask any specific questions. Other situations, like one that occurred with early runs of Eliza, involved people believing that the computer actually understood them and their problem. Some even went through withdrawal when their Virtual Rogerian Therapist was suddenly cut off from them. Does this mean that Eliza achieved true AI? Even it's developers said 'no', that it was only a step towards it.

I think AI should be measured not so much in what a machine knows, but more in what it can learn and apply. This, too, must be more clearly defined. My brother had a video game some years back (How far back? It was on his Amiga...) that actually seemed to learn his tactics. He was playing a German fighter pilot escorting bombers over England. His pet trick was to come in low over the airfield, wobble his wings and put his gear down in surrender. The game would respond by sending two Spitfires to escort him to the ground where he would be expected to land and surrender, and hence lose the match. While almost on the ground, the Spitfires would become lax in their guard, confident that they had won. My brother would then retract his gear, throw a death-defying loop, and waste the Spitfires while straifing everything else around him. The next time he flew that particular plane, they didn't buy his surrender and shot him out of the sky. He flew a different plane under a different name, and the trick worked again. He flew a different plane under his old name and the trick worked again. (Each time, it only worked once. Then, when he flew again, they were ruthless in splashing him.) We ran several variations to see if it was applying the trick only to a certain scenario, but found that the trick remained in memory even when we switched scenarios and played the British. Once the trick was learned, the game would not accept his fake surrenders anywhere on the map, not just at airfields. Once, it even used a variation of the trick against him. He liked to follow his kills down, watching them explode as they hit the ground. The game began to fake fatal dives, let my brother follow in bloodlust, then send a second fighter in behind to spank his tail while the first one pulled out of his dive. None of this game behavior surfaced in the first games, even though he scored high and was an dvanced player. It wasn't until he began tricking the program that it seemed to learn. It was pretty amazing.

Was it AI?

[This message has been edited by Oxygen (edited January 15, 2000).]

 

Log in or Sign up to hide all adverts.

That sounds like a good game--well programmed.

How has AI been applied to the new continuous speech rercognition software? Don't these programs learn and then adjust their data to conform to future input? Could this technique be applied to a larger, more ambitious project.

I'm also curious about the models on which past AI efforts have been designed?

------------------
It's all very large.

 

Log in or Sign up to hide all adverts.

I tried that Dragon Naturally Speaking program. I'm glad I got it a discount because I wasn't impressed. It took a long time to set up and had to have words that began with capital letters as seperate entries, such as "I would like to go." versus "Would you like to go?", where would and Would have to be defined individually. It turned out not to save me that much time when writing letters and reports because of the various forms of syntax that had to be used to get certain effects and punctuation. It also had a hard time recognizing the 'sleep' command, and many of my sideline conversations got entered into the report. Removing them was difficult, so I grabbed the mouse and just started typing. I also had to talk fairly loud, and don't even think about coughing or clearing your throat before trying to get the program into sleep mode.

If I may have a pedestal and the spotlight for a moment here...I type faster than speech recognition software can respond, anyway. Here's the pedestal and spotlight back. Thank you for allowing my harmless bragging.

Please Register or Log in to view the hidden image!

[This message has been edited by Oxygen (edited January 16, 2000).]

[This message has been edited by Oxygen (edited January 16, 2000).]

 

That's good info. Thanks, you may have saved me some money.

 

Oxygen,

I will gladly hand the pedestal and the spotlight to anyone who has a thought to share. Sticks and stones... y' know. Besides, I don't own the pedestal here.

I have tried one of the lesser speech rec. programs and was very unhappy with it. It wasn't continuous speech but rather single word input. I thought it would be interesting to try writing some source code with it. I'm all thumbs on the keyboard.

------------------
It's all very large.

 

Oxygen got it right when he said that AI is defined as when a human can't tell whether he's talking to a human or a machine. It's called the Turing Test and no application has ever passed it completely. It's not difficult to write a program that will be able to answer simple questions and that can even learn vocabulary. However, to do anything more impressive, it needs to implement something like a neural net. Morphological neural nets are being studied nowadays and they seem to be promising. If anyone has a URL or something on morphological neural nets, I'd be interested in that. Thanks.

 

The Turing test is a load of bull. Intelligence lies not only in mimicking some human behavior, but first and foremost in the ability to adapt and survive. That's what the original intelligence (OI, if I may coin a term) evolved for in the first place. That's the definition of true intelligence in my book.

The traditional (symbolic processing) AI and the newfangled reductionist approaches, like neural nets and dynamical systems -- are all alive and kickin'. But you've got to be patient. The brain is by orders of magnitude the most advanced technology humankind has ever known, and reverse-engineering it will not be easy.

As to our ability to build androids, I can offer a quick back of the envelope calculation. Let us assume for a moment that at some point, science has achieved a complete theory of human cognition. All that remains, is to implement an artificial brain. Let us further assume that the fundamental computational unit of the human brain is indeed the neuron. Assume that neurons compute by adding up effects of stimulation from incoming synapses. The excitatory/inhibitory input is passed through some non-linear function before it is integrated in space and time.

There are ~100,000,000 neurons in the human brain. On average, a neuron has on the order of ~10,000 connections to other neurons (possibly including itself, but that's irrelevant at this pont.) The typical operating frequency of an average neuron is 50 Hz (a max. of 50 spikes/sec), but we'll be content with a mere 10 Hz. Assume we require only about 1000 basic operations per synapse to simulate the entire brain (this includes both the actual calculation/data transport code, as well as the simulation infrastructure code. This assumption could be overly optimistic...). Assuming that we use current computer technology (i.e. classical Turing Machines implemented in silicon), how many operations per second would a computer have to be capable of to simulate a complete human brain in real time? Well, let's see:
it's 100,000,000 neurons * 10,000 synapes/neuron * 1000 ops/synapse * 10 Hz = 1*10^16 ops/sec.
That is, our computer would have to be capable of roughly 10,000 trillion operations per second, or 10,000 teraops (10 petaops). Currently, the most advanced computers in existence are capable of performance on the order of 1 teraops. Cutting edge research is moving toward a 1 petaop computer (which will probably be achieved within ~5 years or so.) A further factor of 10 improvement doesn't seem insurmountable. (To give an idea of relative power, the fastest current desktop computers are capable of roughly 0.001 teraops, and even that only for a brief time under ideal conditions; the more typical performance is probably somewhere around 0.0002 teraops. That means a computer to simulate the brain would have to be as powerful as 100 million fastest modern desktops, the kind that aren't even going to be in retail until Spring, put together.)

Notice that my analysis was very crude, and I could be off by one or two orders of magnitude. But what this boils down to, basically, is that by around 2020 (with perhaps +/- 10 years) or so we should definitely have machines capable of singlehandedly simulating a fully functional human brain in real time. Theoretically.

Of course, these machines will cost billions of dollars, tower to the ceiling, occupy acres of space, weigh hundreds of tons, and consume enough power to supply a small city. And naturally, such computers by that time will be utilizing an absolute state-of-the-art in lithographic integrated circuit manufacturing, with processor features shrunk to the very limits allowable by physics and chemistry. To speed up or miniaturize beyond that point, radically new manufacturing methods, revolutionary computational architectures, and circuit construction in three dimensions (as opposed to the current two) will be required.

Not to mention that we will probably have the power to simulate the brain long before we have the requisite knowledge to actually do it.

So... There are two fundamental obstacles to super-intelligent artificial machines. First is, of course, to understand how to build and program them (including detailed, complete understanding of how the brain works). The second obstacle, equally formidable, is to construct the artificial brains in such a way that they are at least comparable in compactness and power consumption to the human brain. But even then, the human brain would be somewhat more advanced, due to its self-repair capabilities and astonishing fault tolerance.

So in conclusion, (and I'm talking here to any and all who read this), if you weren't already in awe of the stuff inside your skull, it's high time for you to learn to appreciate this marvel of your very own. And also probably time to realize that actual real-world androids even so much as equal to us in cognitive power, are probably at least a century away...

Unless, that is, marvellous discoveries in a furious sequence unravel the workings of the brain in a mere couple of decades. And, after all, it may turn out that the functionality of the brain can be implemented much more efficiently than actually simulating the physiologically dictated structure. Maybe, if we discover quickly enough, and are able to improve upon nature enough, we could see real human-equivalent androids within our lifetimes (well, the lifetimes of those of us who are relatively young right now...)

------------------
I am; therefore I think.

[This message has been edited by Boris (edited January 20, 2000).]

 

Boris,

<img src = "http://users.esc.net.au/~nitro/BBoard_member_gifs/bowser_anim.gif">Once again, your contribution is very much appreciated.


If such a machine could be, what would a software or hardware glitch look like? Do you imagine that it would show its self as a mental illness? And how would we debug such a problem?
------------------
It's all very large.

[This message has been edited by Bowser (edited January 21, 2000).]

[This message has been edited by Bowser (edited January 21, 2000).]

 

Bowser,

Good questions, but I don't know the answer. I guess we'll just have to wait and find out

Please Register or Log in to view the hidden image!

------------------
I am; therefore I think.

 

I wonder if the problems associated with such a machine would be exotic, considering the complexity of the system you described.

<img src = "http://users.esc.net.au/~nitro/BBoard_member_gifs/bowser_anim.gif">

------------------
It's all very large.

 

Well, all of this "AI" stuff brings to mind that scene from "WestWorld"(any of you guys remember that movie)where the robots started to "learn" the tactics of gun fighting(sort of like what Oxygen was saying)and began winning the gunfights against the customers. As old as that movie is, it still gives one the creeps about "AI" going haywire.

 

I don't understand. Why do we need to interact with a machine in the same way as a human. Is this our way of being gods on earth. The creators of an intelligence. Or are we pursuing it because we can?????. Don't get me wrong there are alot of applications for "smart" machines, but do we really need them to be like us?, If I want to interact with a human like intelligence I will interact with a human. I just don't understand the need for it I guess.

------------------
work to LIVE...don't live to WORK.

 

Rambler,

There may be many answers to your questions, but this is my own: if you can build it, then you truly understand it. Among AI's ultimate goals is to completely understand the phenomenology and mechanisms of human cognition. And what better way to find out if you've got it right, than to try and build it, and then see if it works?

Of course, other people may give other reasons too.

But beyond all that lies a very deep philosophical trend -- evolution of intelligence. Before us, life on Earth evolved biochemically over ponderous geological time through a fitful process of directed random search. But we can significantly improve upon that strategy, with a directed deliberate design cycle. Ultimately, the question is not whether we should "play God" -- the question is, whether it would be proper for us <u>not</u> to play God. Indeed, our emergence may be the beginning of a new phase shift in the history of life in the Solar system (and for all we know, in the history of life in the entire universe!) Artificial life and artificial evolution may simply be the beginning of a new epoch in the unfolding of the cosmic tale -- similar to the dramatic shifts that occurred in the past from simple chemistry to life, or from single celled to multi-celled organisms. If that is so, then AI pursuits are not merely some labor of conceit -- they are an expression of our very manifest destiny!

------------------
I am; therefore I think.

 

Do you think that the human race is ready?, do we have the ethics to take over evolution?, I believe that at the moment some technology developments are like giving a loaded gun to an infant as a learning toy... genetic engineering is a good example. Sure it be used as a positive tool but it could end up destroying us. It will certainly alter the human race.

------------------
work to LIVE...don't live to WORK.

 

Rambler,

When has the human race ever been ready for anything? When will it ever be? Ready or not, here the future comes -- and it is up to us and our descendants to sink or swim, just as it has always been.

Ethical implications... What is "ethics"? Can you separate it from religion? Once you do, what is there left for technology to 'violate'? Our freedoms and our sense of equality and justice need not be subverted by any paradigm shifts, technological or otherwise. And as for avoiding disasters, the road is quite clear -- be interested, participate, and don't sit on the sidelines crying "woe is me". You can't beat 'em. So join 'em!

------------------
I am; therefore I think.