h a l f b a k e r y
The embarrassing drunkard uncle of invention.
idea:
add, search, annotate, link, view, overview, recent, by name, best, random
meta:
news, help, about, links, report a problem
account:
Browse anonymously,
or get an account
and write.
Login
Create account.
|
|
|
Please log in.
If you're not logged in,
you can see what this page
looks like, but you will
not be able to add anything.
What is artificial intelligence? A narrow interpretation is something that replicates human intelligence. But there is an unimaginably large amount of other phenomena that might be intelligent, just not recognisable as intelligent by human standards.
The Hutter Prize has been set up in an attempt
to solve this problem. The Hutter Prize is based on the theory that efficient compression is closely related to intelligence. The problem with this approach is that you get lots of data compression and not much intelligence.
I propose a slightly different definition: Intelligence is processing information in an efficient manner.
OK, how does this help? It moves away from the data in vs. data out ratio which hasn’t led to AI.
OK, so how do you calculate efficiency? By measuring the amount of heat generated by the computer or neural network.
When a bit of data is erased it dissipates heat (‘reversible computing’ is based on this fact). So I reckon that a computer or neural network that is processing information efficiently will be generate less heat than an inefficient information processor (all other things being equal).
This has the advantage over the Hutter Prize of being able to *dynamically* test for efficiency of processing information. This also has the advantage of being able to test both analogue and digital systems.
This method could be used when developing/training neural networks. This method moves away from the rigid ‘goal based’ neural network training, which is unlikely to develop real intelligence.
The "Hutter prize" website.
http://prize.hutter1.net/
[jutta, Jun 09 2007]
NewScientist article
http://www.newscien...20of%20the%20brain?
Unfortunately a full copy of this article is only available to subscribers. The premise of the article is the brain is intelligent because it minimises free-energy. The proponent of this idea is Karl Friston (do a google search of his publications if interested). [xaviergisz, Jun 03 2008]
Annotation:
|
| |
I'm not convinced that intelligence can be linked to efficiency of data processing - in fact, I'd almost argue that the opposite is the case. |
|
| |
Take the human brain, it uses around 20W - in comparison, your average laptop uses around 15W - and that includes powering the hard-drives, display, cooling fans etc. |
|
| |
An animal brain (e.g. a mouse's) uses far less - but is probably capable of processing a similar amount of input/output - and by your standard, might (at times) appear to be more intelligent. |
|
| |
Now there's probably a vastly disproportionate amount of data processing going on between a brain and your average laptop, but again, the brain performs this using an electro-chemical mechanism, while a PC's processor is likely to be electro-magnetic. I'd venture that the electro-chemical mechanics are less efficient than electro-magnetic bit-switching in the traditional computer. |
|
| |
My point being that the mechanics of intelligence (and hence the power requirements) probably aren't proportional to the level of intelligence. It should be possible to create an intelligence that is purely mechanical (steam-powered if needs be) that would be horrendously inefficient. To put it yet another way, it's my belief that intelligence is a matter of organisation, rather than one of efficiency. |
|
| |
I don't think I can agree with either of you -- I think that intelligence isn't terribly measurable at our current technological stage of development, nor even really very definable. The human brain can come to conclusions without following the logical path, which, as far as I know, is completely inconceivable to a computer... but yeah. |
|
| |
I don't know enough to really opinionate, so disregarding of this anno may be entirely warranted. |
|
| |
Artificial Intelligence is no match for natural stupidity. |
|
| |
I'm not proposing that power=intelligence. I'm saying this might be useful for comparisons between almost identical neural networks. Also, this is more a thought experiment than a ready-to-test theory. |
|
| |
Note: the heat due to the operation of components in an information processor far outweigh the heat dissipated due to erasing information. This thought-experiment is all about the tiny residual amount of heat. |
|
| |
Imagine a black box (i.e. computer, neural network or other information processor) that accepts inputs and produces outputs. |
|
| |
If the input was 10011111 and the output was also 10011111, there would be no erasing of bits of informations, thus no heat dissipation due to erasing of bits. |
|
| |
If the input was 10011111 and the output was 00000000 then there would be 6 (or maybe 8) bits of erasing heat dissipated. |
|
| |
If the input was 10011111 and the black box outputed every second bit i.e. 1011, only 4 bits of erasing heat is dissipated. (this is essentially a very lossy data compression) |
|
| |
So I'm proposing if you had a much more complex black-box and information input, measuring the erasing heat might be useful in developing neural networks. |
|
| |
This is an attempt to approach AI from a physics perspective. |
|
| |
A brain (artificial or real) is like everything else in the universe - it moves on the path of least resistance towards its lowest energy level. So a brain is a form of matter that processes information not because of some innate 'life force', but merely because the path of least resistance is to process information and process it efficiently. The tricky part is designing the brain. |
|
| |
As an equation it's something like:
f(information, brain) = intelligence |
|
| |
Where function f is, of course, incredibly complicated. |
|
| |
bigsleep, I completely agree about the need to hard-wire neural networks to develop AI. See my other idea, "modular neural network" for a possible approach. |
|
| |
What [zen-tom] said, only much louder. Does that make me more intelligent then [Zen-tom]? |
|
| |
It depends [zeno] on how loud you shout, and what you had for breakfast! |
|
| |
//A brain (artificial or real) is like everything else in the universe - it moves on the path of least resistance towards its lowest energy level. So a brain is a form of matter that processes information not because of some innate 'life force', but merely because the path of least resistance is to process information and process it efficiently.// |
|
| |
I like this notion of things moving towards their lowest energy level, taking the path of least resistance - but again, isn't life (and as a high-point of that process, brains in particular) something that prolongs that energy drop via a the twin loopholes of history and organisation in the laws of thermodynamics? |
|
| |
In other words, the shortest and most efficient route for a metal ball to get from the top of the machine to the bottom is directly from top to bottom. A suitably constructed (life analogous) pinball machine stops the ball taking the most efficient route, causing it instead to take a more interesting route, and in doing so, produce the most amount of information possible (in terms of the ball's trajectory) for the same energy drop - - Ahh!, which I suppose is kind of what you are talking about. i.e. Using your measure, a more 'intelligent' pinball machine would be one that made the ball bounce around more interestingly than one who's ball followed a direct path. In these terms, the more the ball bounces about, for a given energy drop, the more efficient the machine. |
|
| |
The problem is defining what it is that's producing the 'information' - e.g. using the last example, you could describe the trajectory of the ball in various ways, all of which might take up the same space, whether it took a straight-route, or a more chaotic one. |
|
| |
Then there's the question of linking this concept to intelligence - complexity certainly and, I suppose if you're in the camp who feel that sufficient complexity, having once reached some critical mass, inevitably becomes intelligence. |
|
| |
But I'm thankful (hence bunnage) to you for making me think about something I've not spent much time on recently - I'm really liking this thing about success being linked to eeking out the most complexity for a given (potential) energy drop - it provide a metric for measuring life - which has, since inception, been improving on pinball designs, getting the ball to bounce around in ever increasingly complex patterns, delaying its inevitable drop back into the drain behind the flippers. |
|
| |
hi zen_tom, thanks for reconsidering this idea. I admit I could have expressed the idea more clearly initially. |
|
| |
I think what we're talking about (your pinball analogy and my artificial intelligence idea) is the general principle that information, complexity, intelligence and entropy are all somehow interelated. The trick is to apply this interelationship to something useful. |
|
| |
/If the input was 10011111 and the output was also 10011111, there would be no erasing of bits of informations, thus no heat dissipation due to erasing of bits./ |
|
| |
If no deletion = less heat dissipation = more efficient processing = more evidence of intelligence, would increasing the output produce even less heat dissipation? If the output was the same as the input, but doubled, this would be less heat dissipation yet. As the output increases, heat dissipation decreases. As output increases towards infinity, evidence for intelligence also increases towards infinity. |
|
| |
hi bungston, you've pointed out an important feature (or is it a flaw) in the experiment. Yes, the output of the black-box must be limited. In fact, in the purest form of the experiment, the black-box would have no ouputs at all (the black-box would have memory and computing components). |
|
| |
Thus the black-box would have the options of:
a) storing the information uncompressed;
b) storing the information compressed (lossless or lossy);
c) erasing the information;
(or a combination of these options) |
|
| |
Of course this raises a really important question: would more heat be dissipated by calculating and compressing than simply erasing data? |
|
| |
Another thing I should emphasise about this idea is that it is a 'calculus of variation'. It is about getting a 'brain' and making small variations and testing the heat output for a particular input. |
|
| |
in maths terms its something like: |
|
| |
minimum erasing heat( f(information, brain+delta1), f(information, brain+delta2), ... ) = maximum intelligence |
|
| |
where delta1, delta2 etc are small variations to the brain, and maximum intelligence is a local maximum. |
|
| |
Personally when in doubt, refer to the
three laws. If anyone didn't get that, go
read some Asimov. |
|
| |
I always thought the brain to body mass ratio was the determining factor. |
|
| |
//I always thought the brain to body mass ratio was the determining factor.// |
|
| |
That belief may require drastic revision. Chihuahuas have the largest brain to body mass ratio out of all dogs... and their brain to body mass ratio is far greater than ours as well. I wouldn't venture to consider them particularly intelligent though. |
|
| |
Everyone's skirting around the key issue here, so I'll just say it: the real trick is not in recognizing intelligence in another form, but self-awareness. Intelligence (IMHO) is tied into reasoning ability, and true reasoning takes self-awareness. |
|
| |
My dogs are self aware: when they look at a mirror, they recognize themselves. If put a spot of paint on my dog's forehead and hold up a mirror, he will look at his reflection and then rub his face with his paw. I performed this experiment after reading about the same thing being done (equally successfully) with dolphins. For the record, my dogs are American Pit Bull Terriers, which are, in spite of thier undeserved notoriety, reknowned as one of the most intelligent breeds. I have also seen my dogs use tools: they occasionally use a stick gripped in the mouth to scratch that hard-to-reach spot on thier butt, and I once observed three of them working together to build a pile of rocks high enough that they could stand on top of it to see over the fence (apparently for the purpose of barking at my neighboors). |
|
| |
My point is that abstract reasoning and goal-oriented behavior are far more reliable signs of intelligence than sheer capacity for logic. I wonder how capable we are to recognize such things in a non-biological entity, or, for that matter, if it would recognize the same traits in us. |
|
| |
I have met many chihuahuas, and not one of them recognizes themself in the mirror. |
|
| |
//If put a spot of paint on my dog's
forehead and hold up a mirror, he will look
at his reflection and then rub his face with
his paw.// I don't believe that but, if it's
true, you could convince me by publishing
it in a peer-reviewed journal. It would be
quite significant. |
|
| |