Machine Sentience

[copiae]

Hrm. This probably isnt the best time to start this thread, but I thought it may be interesting discussion material over the silly season. If people are interested in replying, then they are welcome to, however, I can't guarentee the frequency of my responses for the next three weeks, as I will be overseas.


The thread was originally the second part of a thread I created over at ebla, but the first part is a bit ... irrelevant to the second. However, I've had to edit this post to make it more self-standing... and I've also used that opportunity to fix up a few things. Anyway, without further ado, here 'tis.


I. The Uncanny Valley (http://en.wikipedia.org/wiki/Uncanny_Valley)

The Uncanny Valley is a rather interesting phenomena that occurs within human perception. Basically, if one were to imagine a continuous axis of 'humanness' as an entity goes from being less humanoid to more humanoid, we like it more and more... up to a point. Just before the entity becomes completely human, there is a consistent slump in peoples reactions to it - from empathy to antipathy. This slump continues until full humanness is reached (after which we start liking it lots again).

This slump is also why people react quite adversely to certain computer generated characters (I've found that incorrect motion and/or anatomical disproportionality tend to produce the most antipathic responses in me). I believe that the Uncanny Valley is in itself nothing remarkable... it is representative of a greater human ability: our ability to see dissimilarities. No doubt this ability has served us well in the past - to take a stab in the evolutionary dark, perhaps to be able to spot the difference between cool refreshing shade, and cool refreshing shade containing hidden predators. Currently, an example of its use is to detect outsiders in social groups... Usually some trait gives it away, and once we are alerted, other, smaller disparities immediately become apparent.

I propose the same process is at work in the Uncanny Valley. Up to a point of humanness, entities are extant in thier own right. Once they get past this point, they lose thier extantness, and we begin to compare them with ourselves. Their lackings become apparent, and empathy is replaced with antipathy. This behaviour continues along the axis of humanness, until full humanity is reached. In the past, when it was believed that race and class differences where inherent, its easy to see how the Uncanny Valley could have applied in these situations as well. This well of antipathy, our own intrinsic hatred of that which is not us, has been the fuel that fed the fires of many inequities and deaths in our past.

This hatred is our legacy - whether we choose to discard it or to nurture it with half-truths and fairy tales. It is also this hatred that is the driving force behind mob mentality. If I am correct, and the Uncanny Valley can be applied to other humans too, than what chance do robots stand? Of course, if all we were talking about were humanoid bodies, then I don't think it would be much of an issue - yet - but what if the Uncanny Valley can be applied to humanoid thought?

Unfortunately, there is no real way to verify this - the only animals that come close to aping us are apes, and their thought mechanisms are far too primitive for the Uncanny Valley (if indeed it is present for thought) to kick in. Its not all doom'n'gloom though, just as we overcame beliefs about inherent class and race differences, so too can we overcome beliefs about humans & machines, assuming that we have enough time, and the inclination to.


II. Artificial Intelligence (AI)

At present, there are two types of limitations in programming, systemic limitations (compiler / operating system / CPU / etc) and programmer limitations - There is a adage within the software industry:how smart a program is depends on how smart the programmer is. These days, systemic limitations are rapidly shrinking, as are, to a lesser degree, programmer limitations, as better training, experience and documentation practices produces smarter programmers.

Anyway, any intelligence[1] currently exhibited by a machine is artificial, as to use the corollary of the adage mentioned above, Any intelligence currently displayed by a machine is that of its programmer(s)... We could say that presently machines lack any intrinsic intelligence. There are many theories on how to induce intrinsic intelligence in machines, most of them being algorithm based. [2]

I will hypothesize on what it is that machines presently lack - creativity. A machine cannot extrapolate from existing data in the face of unknown situations; it is incapable of... guessing. It is due to this limitation that logical paradoxes are such effective weapons against machines in Sci-Fi - these paradoxes cause the mind of the machine to enter a livelockesque state[3], or cause the machine to go outside its programmed bounds, and enter an erratic state. The other ability granted to humans by creativity, which is currently lacking in machines, is recursive thinking (thinking about thinking), which allows us humans to neatly sidestep these paradoxes - we gain ...control (using that term very loosely) over our thought processes.


So how do we induce creativity in an Artilect[4], to test if my hypothesis is correct? I havent the faintest idea.


III. The futures overrated.


Lets assume that we can somehow induce AI. Why do we want to? "To improve the quality of human life"? "To see if we can"? What role will AI play in the future? Will it be our slave? our servant? our equal? our superior?

We can perhaps look towards science fiction as an answer to these questions. Science Fiction paints intelligent machines with three very broad brushes: benevolent, neutral and malignant. Benevolent AI is completely subservient (the Jetsons), neutral AI may have its own agenda, but nonetheless are usually subservient (Star Trek/Wars), and malignant AI usually start off as being subservient, but when threatened with extinction, react, or develop megalomaniacal tendencies (Terminator).

There is a pattern here. When thinking of AI, we tend to lump it in with 'intelligent technology' - any embedded system (things like mobile phones and microwaves). These things are inherently subservient, and so that property gets passed on to Artilects as well. Subservience implies control, something which isnt a problem in mobile phones and microwaves, as they possess no will to power. Based on what I see on the internet, when these concerns are raised with scientists, they are usually batted away with a wave of the hand, and a comment along the lines of "even if we could induce AI tomorrow, it will be a long time before that becomes anything close to an issue", or a vague comment of something along the lines of the three laws of robotics taking care of that.

Isaac Asimov, in his celebrated science fiction novels, developed the three laws of robotics:

1. A robot may not harm a human being, or, through inaction, allow a human being to come to harm.
2. A robot must obey the orders given to it by human beings, except where such orders would conflict with the First Law.
3. A robot must protect its own existence, as long as such protection does not conflict with the First or Second Law.

Once more, we see the mastery that humans have over robots. The robot is subservient to us, it must serve us, even if it costs it its own life. As far as I can see, there has been little work done on the ethical considerations of this - in essence, we are creating intelligent slaves for us. Can this be justified? If we were to stick a human in the machines place, clearly it would not. Indeed, human myth is rife with tales of cruel tyrants overthrown by heroes, and sacrifices made for freedom.

Speaking of freedom, now we come to the flip side, control. For humans, there are two different mechanisms for control: self-control and control of others. Self-control is the control of ones desires, usually for a higher-order purpose (e.g. I do not constantly eat junk food, lest I get fat/fatter.) Control of others, however transitory it may be, is usually achieved through emotion (primarily fear and love), or the threat of emotion.

How can we control an Artilect? Perhaps emotions come bundled with sentience, but it seems unlikely to me. (no reason, just my opinion). Anyway, if I am correct, and creativity is the key to unlocking sentience, then the laws of Asimov (or any other non self-preservation law), become about as meaningful as the ten commandments, or any other laws that humans can live by. Creative machines gain the ability of recursive thinking, and thus gain a form of control over thier thoughts - ah, but some would say, surely the laws would be hard coded - no matter how well the machines could think, the control would be more analogous to the movement of your elbow - try as you might, you simply cannot extend your arm beyond 180 degrees. On the contrary, you can, though you may have to use an external device to do it, and the need has to be enough to overcome pain avoidance (fear). An Artilect feels no pain - the first time it does this, it will realise the absence of repercussions, and a law without some form of repercussion is not worth the paper its printed on. If I am wrong, and creativity has nothing to do with sentience, I would still consider it reasonably safe to assume that an AI has the ability of recursive thought.

Sadly, I don't think many people have spent much time on considering control (and if it is possible) and the ethical considerations of controlling an Artilect. If it somehow gains the ability to write itself, there is absolutely no way we can control it, short of physical isolation. We are creatures of flesh and bone, and though we created the digital realm, we are effectively outsiders in it. Imagine an intellect for whom creating code is about as easy as for me to flex my arm, or to wiggle my fingers. There would be no digital security system capable of containing it, and no human programmer capable of matching it. Further, if it learnt how to replicate its actual sentience, it could ...breed... by sending out a virus that is essentially a seed of itself, to fill the shell (another computer).

A lot of people consider us to be safe from a megalomaniacal Artilect/Artilect fighting for its life by simply not connecting it up to any sensitive military computers, but unless these computers are totally isolated (i.e. have no physical connections with any other computers), they can be easily accessed by any Artilect knowing what it is looking for. It is also very possible to create a global holocaust without access to these computers. How?

Well, lets see... If it is a clever Artilect that is patient, and the public does not know of its existence, it can cause events to occur that closely resemble terrorist actions, and create a trail of clues that point to a rival super-power, and act as a catalyst for a global world war. If it is not patient, It can gain access to NASA's computers, and instruct the hundreds of satellites in the sky to gather speed where possible, and come raining down on us - a fiery apocalypse? . Worst of all, in the situation of a rogue Artilect[5], it becomes like a ghost in the shell: impossible to find, and impossible to kill, short of killing its shell - Destroying all of our advanced technology (all computers, communication methods and embedded systems).

Some food for thought. Thanks for reading it =)





[1] I'm defining intelligence as applied pattern recognition (creativity). I'd be happy to discuss this definition in another thread, later on.


[2] There are two types of algorithm-based AI efforts... Strong AI and Weak AI. Weak AI research looks at making more clever algorithms, and Strong AI research deals with inducing AI. In here, I am dealing exclusively with strong AI.


[3] Livelock can kinda be visualised as something being stuck in a catch-22 loop.


[4] an artificial intellect (used to represent a hypothetical AI entity). Term borrowed from here (http://www.cs.usu.edu/~degaris/artilectwar2.html).

[5] One that is no longer chained to our will.

[JoeP]

Interesting post. I hadn't heard of the uncanny valley; sounds reasonable but I suppose the truth behind it will be more complex - common sense that it applies in some cases, but will apply differently or not at all in different cases.

Your comments in II and the start of III are fair. Nitpick:
An Artilect feels no pain
How do you know? We haven't constructed an artilect yet. Who says pain or some analogue of it isn't essential for the functioning of a self-sufficient, surviving sentient entity?

The latter part of your post, however, seems to fall into a common (imo) trap of dwelling on those sf archetypes. I think this misses a huge area of possibility - and of current valid debate - of artilects who are beyond 'neutral'. Being malignant is not the only outcome. Have we progressed at all from '50s sf?

Some dimensions I'd like to see considered:

What would be the (psychological) nature of an artilect created as an "adult", ie not going through an infancy/childhood/adolescence development? We have no models of such a sentient being to suggest whether it could be much like us.

What would be the nature of an artilect created as a one-off, ie not part of a community? Same comment applies.

You're talking about a program that could travel around cyberspace and defend itself. We can't even get the same program to run on a Mac as on Windows mostly; let alone trying to reimage a complete working system on a hardware platform with a different drive configuration. So I think this is far-fetched and therefore far from being realistically understood (OK, it's a matter of technical limitations which can be overcome, as you addressed in your post).

And the core topic is this: if something sentient is capable of self-defence and malignance, isn't it also capable of other stuff: art? creating unique products? trade and commerce? building stuff? And hey ... creating artilects more powerful than itself ...

[Adora]

Unfortunately, there is no real way to verify this

I was under the impression some Japanese companies are running these kinds of audience-reception tests in regards to virtual idols and robots.

But that may have just been a rumour.

PS: Please leave Asimov at the door. X_X His books may be "celebrated", but his laws suck and are counter-productive to AI (or artificial human intelligence) evolution. You need to divide a line between "robot" and "AI". A "robot" is exactly what its name means - a subservient worker machine. AI cannot be that.

[Farren]

I have to pitch in and agree that Asimov's "Three Laws of Robotics" are simplistic, almost childish ideas and his "Robot" books are painfully contrived and unlikely.

Also I'd like to say speculating that creativity is the "missing ingredient" in attempting to achieve human like sentience seems a little fluffy to me. A serious examination of computers and their possibilities reveal some more glaringly obvious problems:

1. Most current technology is digital. Digital computation of necessity can only approximate an analog phenomenon. No matter how apparently good the approximation, it will always be infinitely different. Whether this difference is trivial or non-trivial is debatable but it seems reasonable to assume its a non-trivial difference. That said, we aren't precluded from building analog computers.

2. Human sentience is underpinned by massive parallelism not reflected in current computing technology.

3. On a tangentially related point, human brains appear to be much faster than current computers. Expert estimates float around the 17 tera-flop (floating point operatons per second) range. This is by virtue of parallelism. Signal transfer in the brain is actually substantially slower than modern electronics - but far more signals are being processed at any given time.

[Farren]

The uncanny valley is a disputed concept from a scientific perspective, as indicated in the linked article. I've only encountered it in computer animation. I wasn't aware the idea originated in robotics. Certainly its gained some traction in the former field because stylised computer renders do seem more pleasing to the eye than almost-but-not-quite real looking stuff.

Personally I don't think the same instincts which motivate racism etc inform the discomfort felt when viewing almost but not quite perfect rendering. The former relies heavily on our tendency to cluster percieved groupings, as a shortcut to suitable responses when dealing with other human beings. Its a combination of our pattern recognition and social response mechanisms. From examination of my own reactions to different computer animation I think the Uncanny Valley in computer animated images is driven by a simpler process.

Stylised stuff cuts out the in-between stuff and emphasises the visual features our brain quickly responds to (like our programmed response to big eyes, presumably due to nurturing instincts), soliciting a pleasing ease of interpretation in the process. Almost-but-not-quite-real stuff is as subtle and requires as much mental effort as reality (like ambiguous expressions) but with an unnatural element (the less than perfect rendering that makes it not quite real) that unnerves at the same time. The former is easy to interpret, perhaps easier than reality (I know I often recognise a charicature of a famous person quicker than a photo at an odd angle). The latter is just unnerving.

[Adora]

There's a context you have to take in regarding the Uncanny Valley in regards to computer animation. Because it is a form of animation. It's entertainment. It's a fantasy. It shouldn't be trying to be too real.

This is where you get your problems with Polar Express and Final Fantasy: The Spirits Within. Polar Express personally creeped me the fuck out because I don't like Tom Hanks, and it was like a movie of millions of cloned Tom Hanks's. Utterly. Terrifying. But yes, you definitely got the Valley going on there.

But the major problem we have with CG is that it tries to hard to be real, but no originally real. It uses actor's faces far too often to model the characters on, and more often it is these movies that do the worst. They play off the real rather than off the fantasy, whilst still abusing the fantasy.

You got this again in Final Fantasy: TSW. If I wanted to watch Steve Buscemi play "kooky SF soldier" mode, I'll go see a movie where he actually does that. I don't want a motion-captured, voiced and modelled version of him doing it, simply because they can, which is so often the case with CG.

Like I said- it tries so hard, but it is so far from. They can map the basic shallow movements, but technology is not advanced enough to simulate light refraction through skin, or through hair, or eyes, and even complicated programmes have to be created to simulate different materials, ala Monsters Inc (a particularly fine example of a successful animated CG movie). We can't create believable synthetic voices or speech patterns from scratch, and motion capture is still majorly relied upon for characters' movements in the films. It can't come into philosophies of "The Other", because it is not yet equal enough to be "The Other" as it should be.

[Farren]

I watched a Playstation 2 intro movie the other day that used motion capture to good effect. It was an epic battle, with lots of whole people but very few facial close-ups. What little talking scenes there were often had the near character shown from behind and the other character facing the viewer at a distance.

It was astonishingly real, so much so that only the subject matter hinted at the fact that it was CG. But they did have one or two bits with close-up faces and those had the unnerving quality we're talking about.

I think our faculty for facial recognition and interpretation is immensely more developed than the way we process the rest of the body. I read or saw something recently that said an average person can remember something ridiculous like 400,000 individual faces. The subtlety of discernment facilitating that must be incredibly fine tuned.

From my own extensive experimentation with modelling and games programming I think a hug part of it is in the way the face moves. I've specifically experimented to try understand what makes faces look so damn "plastic". I bought a games engine earlier this year and at one point tried to model a friend. I took digital photies of him, mapped the textures to a high-polygon-count head then tweaked it to get it looking exactly like him in the default facial posture.

It looked remarkably lifelike but as soon as I started animating it to pronounce different phonemes it didn't look like him. Just the fact that each side of the face is controlled by a different side of the brain means the way the mouth forms an "o" will be subtley lopsided in a specific way for each person. More, it will be differently shaped depending on what expression immediately preceded it.

The phonemes were created using skeletal animation - inserted invisible "bones" under the skin and specifying the degree of influence each "bone" had on the polygons of the mesh, then moving/resizing the bones to get a corresponding mesh distortion. Of course the way I was using it the "bones" were more like "muscles". The advantage of this is than the same facial expressions can be applied to multiple different models just by importing the bones and animations from one model to the next.

Remodeling the actual individual model for every expression by hand (i.e. modelling what he actually looks like saying "o" and so on) worked a lot better, but then the face snapped from one expression to the next without transitions. The result is that the model is quite photorealistic, but the animation is effectively still "stylised" in a sense, snapping from one archetypal pose to another.

The elastic nature of skin over muscle and bone also produces additional movement. Lips are soft and bulge and flatten in specific ways in response to how they are moved by muscle. I don't think simulating these to a degree unrecognisable as artifical is impossible. I just think its going to require significantly more detail and computing power before its as realistic as some of the body animation I've seen is.

Back on the OP, I think it is an important consideration if one is thinking about human-like AI. A large part of our learning and appropriate response mechanism revolves around how other people respond to us. Until a machine looks human, its going to be proscribed at least in some ways from understanding humans in the manner we generally call empathy, since humans will respond differently to it from the way they respond to each other.

Simulating a human face with machines, though, might curiously enough be easier than simulating it using CG (maybe), because asynchronous analog facial circuitry, correct anatomical modellng of muscles and suitably elastic and spongy flesh would likely produce many of the same unintential artifacts of motion that real facial movement does. By unintentional artifacts I mean those not consciously willed by the brain, like the flattenings and bulgings of flesh when you say "o".

[Clutch Munny]

The eyes.

I saw FF:TSW in the theatre, and saw all the evidence of the much-hyped jillions of dollars spent on getting the hair motion perfect. But guess what? We don't read intentions and emotions off hair motions. Eyes, on the other hand, are crucial, and the CGI eye-motions were, as peer aptly notes, right in that valley of close enough to be a human reaction but not good enough to be the right human reaction.

So, strangely, the overall effect was of intelligent agents doing really bad acting: they didn't look at each other right, so they seemed just to be mouthing their lines.

[justaman]

Eyes, on the other hand, are crucial, and the CGI eye-motions were, as peer aptly notes, right in that valley of close enough to be a human reaction but not good enough to be the right human reaction.
I had been personally struck by the way Gollum was portrayed in the Two Towers when he is reminded that his name is 'Smeagol'. Watching the making of documentary, they note the same thing, going so far as to call it one of the most poignant acting moments in the whole movie, and I don't much disagree with them, it was very good. And I think it was precisely because of the small muscle twitches in the face that was so perfectly human - because it was so pain-stakingly copied from Andy Serkis' portrayal. It wasn't like Jar-Jar spewing nonsense lines, it was a CG character acting in a believable, humanoid manner.

Having said that, though, I'm not sure CG Gollum avoids Adora's criticisms, but I still thought the end result was very good. :yup:

A little more on-topic, I think AI ethics would have to become a very important field, though I think we've some way to go before we really have to start concerning ourselves about it.

I'm fairly sure that the biggest difference between computers and humans which makes programming AI difficult is the human ability to spawn, remove and regenerate hardware, where as computers are stuck with what infra-structure they are born with. Humans are constantly building new neural networking chains - analogous to transistor series in digital circuitry of course - which is a feat I'm not sure we'll ever be able to emulate.

Humans also have the advantage of having many to many connections for each individual neuron - as many as 150,000 - which is something we simply cannot do with classical digital circuitry, even if humans have numerically fewer neurons than modern computers have transistors.

But I think it'll eventually happen, especially with things like quantum and optical computing making some headway. I think the OP makes a lot of very good points. It would be very difficult, I think, to hardwire something as restrictive as 'don't hurt humans' into an artillect (I like this term!) and yet still keep its ability to be creative unfettered.

I also think that if a computer ever voiced a desire to suicide, that would probably be a pretty good indication that we're making headway :P