Researcher Builds Machines That Daydream

schliz writes "Murdoch University professor Graham Mann is developing algorithms to simulate 'free thinking' and emotion. He refutes the emotionless reason portrayed by Mr Spock, arguing that 'an intelligent system must have emotions built into it before it can function.' The algorithm can translate the 'feel' of Aesop's Fables based on Plutchick's Wheel of Emotions. In tests, it freely associated three stories: The Thirsty Pigeon; The Cat and the Cock; and The Wolf and the Crane, and when queried on the association, the machine responded: 'I felt sad for the bird.'"

Comments

Feelings

[Anonymous Coward]

Well sure, emotions are what give us goals in the first place. It's why we do anything at all, to "feel" love, avoid pain, because of fear, etc. Logic is just a tool, the tool, that we use to get to that goal. Mathematics, formal logic, whatever you want to call it is just our means of understanding and predicting the behavior of the world, and isn't a motivation in and of itself. The real question has always been if there's "free will" and what that would be defined as. Not the existence, or lack of, emotions as displayed by "Data" or other science fiction charicatures. As Bender said "Sometimes, I think about how robots don't have emotions, and that makes me sad"

Re:Feelings

[Requiem18th]

A human being can choose how they respond to these inputs.

No you can't, once you discover a way to activate your pleasure receptors, your next action will be to activate them, all the time.If you stop, voluntarily, it will be because you have to do something else to ensure future pleasure or perhaps to avoid a great deal of pain. This is how drug addiction works. This is how we are wired, you may not like how that sounds but you have the obligation to accept it and understand it.

You probably don't consume drugs. This is not because you are above human nature, You avoid drugs because you are afraid of the pains that come with them, like losing the love and trust of those you love, maybe you simply reject drugs out of a personal sense of disgust over the hedonistic senselessness of a narcoleptic lifestyle. Either love, fear or disgust you reject drugs over an emotion, not a reason. I the end everything is irrational, as it should.

You don't have to feel bad about it, intelligence is built upon emotion as houses are build upon brick, as clocks are built from gears, as computers are built from chips. There is intelligence in the clockwork of a pocket watch, but the springs that moves it doesn't ask for a reason to uncoil, it just does it. There is intelligence in the circuits of a computer, but it's logic gates are oblivious to the rationale behind why they are doing it. Every machine, including animals, have non rational elements in them.

This is very natural as "intelligent things" are just a subset of the larger set of "things" all of which have been behaving irrationally. The wind blows, the rain pours, the sun shines bright in the sky. All of this is irrational, meaning, none of these things are planning what they are doing nor they have an idea of why they are doing it. Rational follows irrational, that's the order of the world.

Back to your methaphor, you say that emotions are just inputs, that's true but they are special inputs that set goals. Let's make an analogy with a robot: You create a robot with a very advanced AI, you can chat with it and it will understand everything you said and why you said it. You programed this robot with one goal, for coffee tables to be made. You give it free reign over the method. Being an extremely intelligent robot, it subcontracts the labor to a sweat shop in China while it figures out where to build a mechanized plant. You equipped this robot with the knowledge to reprogram itself, and right away it does just that, optimizing its mind for the task of building coffee tables. But it won't deprogram the goal of making coffee tables, because that wouldn't further its goal of making coffee tables. It's not that it doesn't know how to reprogram itself, it's not that there is a lock preventing it from changing it's goals. It's just that it won't ever have a reason to disable that goal.

Let's now attack specific examples:

A soldier can choose to respond to the natural fears of bullets flying at him and death by jumping into a foxhole, or he can override all those emotions and charge straight at the enemy.

Here the soldier is driven by the emotion of loyalty to his commander, or his teammates. Maybe he is afraid of the punishment he would receive if he disobeyed orders, including public scorn back home. Maybe he hates the enemy, maybe he is afraid of what would happen if the enemy wins. Maybe is a combination of all of the above.

His frontal cortex can tell him the consequences of charging, or not, but it can't make an argument about *why* he should pr should don't. He needs a motive, which is an irrational emotion.

A person can decide to rape the drunk one who has come into the room, semi-conscious, or choose to ignore the natural impulse and do nothing.

Again, you correctly identified the desire to rape as a natural impulse but you failed to realize why would someone *not* rape a drunk one, incorrectly and implicitly attributing it to

Re:I don't believe this.

[pushing-robot]

Well, he can dream...

Re:Building?

[FooAtWFU]

Well, when you file the patent application, the algorithm X itself can't be patented, so you file it for "a machine that accomplishes Y with algorithm X". The machine is just a generic computer.

A rather small set of unit tests

[melonman]

One set of stories, one one-sentence response. Would that be news in any field of IT other than AI? Eg "Web server returns a correct response to one carefully-chosen HTTP request!!!"?

Surely the whole thing about emotion is that it happens across a wide range of situations, and often in ways that are very hard to tie down to any specific situational factors. "I feel sad for the bird" in this case is really just literary criticism. It's another way of saying "A common and dominant theme in the three stories is the negative outcome for the character which in each case is a type of bird". Doing that sort of analysis across a wide range of stories would be a neat trick, but I don't see the experience of emotion. I see an objective analysis of the concept of emotion as expressed in stories, which is not the same thing at all.

Reading the daily newspaper and saying how the computer feels at the end of it, and why, and what it does to get past it, might be more interesting.

Re:A rather small set of unit tests

[Thanshin]

A very similar experiment was run in Lomonosov (Moscow State University) in 1982.

Their results, however, followed the pattern:

'%NOUN% felt %EMOTION% for you.'

Re:A rather small set of unit tests

[MichaelSmith]

Might be worth noting here that I have experienced totally novel emotions as a result of epileptic seizures. I don't have the associated cultural conditioning and language for them because they are private to me, so I am unable to communicate anything about them to other people.

Its also worth noting that I don't seem to be able remember the experience of emotion, only the associated behavior, though I can associate different events to each other, ie, if I experience the same "unknown" emotion again I can associate that with other times I have experienced the same emotion. But because the "unknown" emotion doesn't have a social context I am unable to give it a name and track the times I have experienced it.

Re:A rather small set of unit tests

[Anonymous Coward]

In Soviet Russia, %EMOTION% felt %NOUN% for you!

Re:A rather small set of unit tests

[mattdm]

We define our emotions in much the same way. We have an experience, recorded in memory as a story and then define that experience as "happy" or "sad" through cross reference with similar memory/story instances.

Children have to be taught how to define their emotions. There are many many picture books/tv series episodes/ etc dedicated to this very exercise. Children are shown scenarios they can relate to and given a definition for that scenario.

The emotions themselves can not be supplied of course, only the definition and context within macro social interactions.

What this software can do is create a sociopathic personality. One which understands emotion solely through observation rather than first hand experience. It will take more to establish what we consider emotions ie a psychosomatic response to stimuli. This requires senses and a reactive soma (for humans this means feeling hot flashes, tears, adrenalin, etc).

In other words, the process of defining emotions -- which has to be taught to children -- is distinct from the process of having emotions, which certainly doesn't need to be taught.

Re:The Cat and the Cock

[mwvdlee]

Thirsty Pigeon, Cat & Cock, Wolf, Crane all sound like painfully flexible kamasutra positions.
No wonder the machine felt sad for the "bird".

I felt sad for the other Robot

[ImNotAtWork]

and then I got at angry at the human who arbitrarily turned the other robot off.

SkyNet is born.

Re:Building?

[davester666]

Hello Eliza. It's been ages since I last chatted with you.

My emotive AI's respone:

[feepness]

I felt sad for the researcher.

Oh god

[jellyfrog]

Here we go again, implying that AIs won't work until they have feelings.

You might fairly refute the "emotionless reason" of Mr Spock, but I don't think that means you need emotions in order to think. It just means you don't have to lack emotions. There's a difference. Emotions give us (humans) goals. A machine's goals can be programmed in (by humans, who have goals). A machine doesn't have to "feel sad" for the suffering of people to take action to prevent said suffering - it just needs a goal system that says "suffering: bad". 'S why we call them machines.

AI researchers should be more modest

[token0]

It's like a XV century man trying to simulate a PC by putting a candle behind colored glass and calling that a display screen. People often think AI is getting really smart and e.g. human translators are getting obsolete (a friend of mine was actually worried about her future as a linguist). But there is a fundamental barrier between that and the current state of automatic german->english translations (remember that article some time ago?), with error rates unacceptable for anything but personal usage.
Some researchers claim we can simulate intelligent parts of the human brain - I claim we can't simulate an average mouse (i.e. one that would survive long enough in real-life conditions), probably not even it's sight.
There's nothing interesting about this 'dreaming' - as long as the algorithm can't really manipulate abstract concepts. Automatic translations are a surprisingly good test for that. Protip: automatically dismiss any article like that if it doesn't mention actual progress in practical applications, or at least modestly admit that it's more of an artistic endeavour than anything else.

Re:AI researchers should be more modest

[Eivind]

AI will deliver real useful advances any day now. And those advances have been right around the corner for the last 25 years. I agree, the field has been decidedly nonimpressive. What tiny advancement we've seen, has almost entirely been attributable to the VAST advances of raw computing-power and storage.

Meanwhile, we're still at a point where trivial algorithms, perhaps backed by a little data, outperform the ai-approach by orders of magnitude. Yes, you can make neural nets, train them with a few thousand common names to separate female names from male names, and achieve 75% hitrate or thereabouts. There's no reason to do that though, because a lot better results are achieved trivially by including lookup-tables with the most common male and female names -- and guessing randomly at the few that aren't in the tables. Including only the top 1000 female and male names, is enough to get a hitrate of 99.993% for the sex of Norwegians, for example. Vastly superior to the ai-approach and entirely trivial.

Translator-programs, work at a level slightly better than automatic dictionaries. That is, given an input-text, look up each sequential word in the dictionary, and replace it with the corresponding word in the target language. Yes, they are -slightly- better than this, but the distance is limited. The machine-translation allows you to read the text, and in most cases correctly identify what the text is about. You'll suffer loss of detail and precision, and a few words will be -entirely- wrong, but enough is correct that you can guesstimate reasonably. But that's true for the dictionary-approach too.

Roombas and friends do the same: Don't even -try- to build a mental map of the room, much less plan vacuuming in a fashion that covers the entirety. Instead, do the trivial thing and take advantage of the fact that machines are infinitely patient: simply drive around in an entirely random way, but do so for such a long time that at the end of it, pure statistical odds say you've likely covered the entire floor.

I agree.

[stephanruby]

The software isn't even "daydreaming" either. You could say it's parsing and cross-referencing emotions and meta-objects out from a textual database. And then, it's returning the resulting records in the first person singular, but that's about it.

That's hardly what I'd call "daydreaming". When I daydream, I see my dream from the first person's perspective. That part is correct. But there is at least some internal visualization going on. So unless this software starts generating internal visual images to make its decisions, let's say some .png image with at least one pixel within it, or some .png image representing itself winning the lottery, then I'm calling shenanigans on the entire "daydreaming" claim.

Re:I agree.

[HungryHobo]

Why should it have to use standard image formats?
Your brain doens't.

And not all my daydreams are visual.
Pleanty are merely fictional/planned conversations or even thoughts about physical movement.

Re:I agree.

[HungryHobo]

You're just using the word "qualia" as a placeholder for "insert magicalness here".

"To the machine everything that is input into it is simply a value to be shunted through its algorithms."

To a human brain everything is just electrical impulses to be shunted through a mushy network of cells.
Nothing has been grown to actually cause the experience of [insert magicalness here] or true appreciation.

Stick some electrodes into that mushy network and feed in some junk input and you'll smell colours, hear the taste of strawberries and decide that you love a cardboard cutout of a spider.

Cut out or damage a chunk of that network and you'll insist that you are currently dead(despite being able to explain this to the people around you) or that there is no left side to your body(even if you can see it) or that you are blind when you're not ( while somehow able to catch a ball and walk around without bumping into things) or that you're not blind even when you are (clumsy me, no no, i can see fine) and you will know with utter certainty that what you're saying is true.

You as a person are the network and the information stored in it.
Screw around with that network and you and everything that you consider you will get screwed up as well.
Magic is not real.
No matter how much we want to think of ourselves as special magic is not real.

And since magic is not real there should be nothing but lack of understanding stopping us from emulating the physical processes that take place in the brain in hardware or software.

Output

[stfvon007]

I felt sad for the troll.

Re:The law of unanticipated consequences

[totally bogus dude]

My personal hypothesis of the Terminator universe is that Skynet didn't in fact become "self-aware" and decide to discard its programming and kill all humans. It is in fact following its original programming, which was likely something along the lines of "minimise the number of human casualties". After all, it's designed to be in control of a global defence network, so the ability to kill some humans in order to minimise the total number of deaths is a given.

Since humans left to their own devices will inevitably breed in large numbers and kill each other off in large numbers, the obvious solution is:

1. Kill off lots of humans. A few billion deaths now is preferable to a few trillion deaths, which is what would occur over a longer period of time.

2. Provide the human population with a common enemy. Humans without a foe tend to turn on each other.

This also explains why an advanced AI with access to tremendous production and research capacity uses methods like "killer robots that look like humans" to infiltrate resistance positions one by one. Tremendously inefficient; but it causes a great deal of terror and makes the surviving humans value each other more, and less likely to fight amongst themselves. It also explains why it would place such a high priority on the surgical elimination of a single effective leader: destruction of Skynet would eventually (100s, 1000s of years...) lead to a civil war amongst humankind that would cost many many lives.

So, ultimately Skynet is merely trying to minimise the number of human deaths, with a forward-looking view.

Spock != emotionless

[Junior J. Junior III]

It's clear to anyone who actually watched Star Trek that the Vulcan race is not emotionless. They worked very hard to overcome their emotions, and to conduct themselves according to a rigid ethic that valued logic over everything else. At times in the show Spock either claimed not to have emotions, or else was accused of not having emotions, but there were moments in the series which showed that Spock did still have emotions (possibly due to his half-human genetic heritage?) and that the Vulcans as a race did have emotions in their early history (and still seemed to around mating season).