It is currently Wed Sep 20, 2017 11:56 pm

All times are UTC - 8 hours [ DST ]




Post new topic Reply to topic  [ 6 posts ] 
Author Message
Offline
 Post subject: AG Explanation for Dummies?
Post #1 Posted: Tue Feb 21, 2017 7:47 am 
Lives in gote

Posts: 575
Liked others: 4
Was liked: 28
KGS: 3D
Tygem: 5D
I am looking for the most intuitive explanation of how AG works, so where u need not too much to know about AI, but still get some deeper inside. So many sources out there, YT, pdf's, any recommendations?

Top
 Profile  
 
Offline
 Post subject: Re: AG Explanation for Dummies?
Post #2 Posted: Tue Feb 21, 2017 8:45 am 
Lives with ko
User avatar

Posts: 214
Liked others: 8
Was liked: 27
Rank: NR
This message is written to be intelligible to young children about how
Alphago's mind works.

First, let's see how your own mind works, so we can compare the two.

Imagine a schoolyard of tweeting children passing little messages
around from hand to hand. Each individual tweets to several friends,
who in turn tweet to others. Tweets flow around the yard like currents
flow around the sea, which a bird's-eye view through a "tweet camera"
could see like water flowing in streams, pools and rivers.

Because of gravity, real water only flows downhill, but tweet water
can flow round and round in circles and spirals and all kinds of
shapes.

Your brain is a schoolyard of tweeting neurons, each of which can have
hundreds or thousands of hands for receiving tweets from others and
passing on tweets of its own. A bird's-eye view of all this activity
seen through the lens of a magnetic resonance imaging camera produces
a video that looks like a city seen at night from high above, with
rivers of car headlights flowing around and building lights twinkling
on and off, as wonderfully filmed in timelapse in the cinematic
masterpiece "Koyaanisqatsi".

The messages that neurons tweet to each other are physically embodied
in concentrations of neurotransmitters with fancy names like
acetocholine and dopamine, but their meanings are simple yells: the
louder one neuron yells to another, the more likely the other will
hear it. Neurons have two different kinds of receiving hands, some
that are turned on by yells, and some that are turned off.

Each neuron receives yells from hundreds, sometimes thousands, of
others at a time. Chemical yells in synapses make waves of ionisations
in receiving hands called dendrites, which flow up to the Chief
Executive Officer (CEO) of the neuron, which is called the axon
hillock because from the outside it looks like a lump in your arm.

Like all CEOs, the axon hillock doesn't do much, because all the hard
work is done by others. All a CEO has to do is every now and then
decide between a shortlist of choices provided by his advisors, and
then tell other people what to do, and so on down the line until you
get to the true value-adders of any corporation or civilisation, the
manual labourers such as computer programmers, bricklayers, and
dustmen.

All a neuron's CEO has to do is every now and then do a simple sum of
all the 'on' yells minus all the 'off' yells and see whether that sum
is big enough to tickle its fancy. If it does, it lights up an
electrochemical wave down its trunk (its axon), which branches out to
its numerous tweet-sending hands (called axon terminals), which pass
its tweet onto those of its neighbours that it talks to.

A neuron's tweet is just one letter long. It's either on, or it's off.
It's binary, just like the signals flowing around inside a digital
computer. Neurobiologists call the tweet an "axon spike all or nothing
response" because they noticed that its intensity and frequency does
not change from one tweet to another - it's either there, or it's not
there..

Now, you may think that just going tweet or not isn't enough to say
anything much, but Samuel Morse knew better, because any letter or
number can be represented as a string of on-or-off tweets (called
"bits" in IT jargon, short for "binary digit"). The code Morse devised
for flashing messages from one ship to another enabled ship captains
to coordinate their activities and fight battles as a team.

Sending messages is one thing, but being able to understand them is
quite another. It's easy to see how messages can be written in binary,
but what about figuring out their meanings? And deciding what to do
about them? How does your brain think? And how do you feel?

If you read back a little bit, you will see that i have already told
you the answer to all these deep philosphical and spiritual questions.

Remember the neuron CEO? It makes a decision; it chooses between 'yes'
and 'no' - between 'on' and 'off'.

Put a few building bricks together in the right way and you get a
house. Or a car. Or a computer.

That's right, your brain is a biological digital computer made of
billions of neuron CEOs and their assistant dendrites and axopn
terminals, assisted by manual workers like glial cells which keep
their bosses well fed and clean.

And so, finally, we come to Alphago.

She isn't a biological digital computer, nor is she just an adding
machine like a pocket calculator.

Or is she?

Alpha has lots of neurons too; they're a bit different from biological
neurons but not all that different, in terms of the function they
perform. They have their own electronic equivalent of tweet receving
and sending hands and their own equivalent of the neuron CEO which
does a pretty similar job - because its function was inspired by the
function of the biological neurons that are found in all animals with
brains, including sea slugs and starfish.

But that's where the similarity ends, for whereas biological neurons
can grow and shrink, and live and die, and make new friends and lose
old ones, artificial neurons are more like soldiers lined up in rows
that only do what they're told to do when they're told to do it.

Nevertheless, they can be told to learn, to change their CEOs'
decisonmaking behaviours. If all goes well, CEOs are given a bonus,
but if things go badly, they are turned down - if only that happened
to the human CEOs that robbed the poor of their savings in 1929 and
2008 and did all sorts of other awful things like starting and
financing wars! But that''s another story....

One drop in the ocean doesn't make much difference, and changing one
CEO's mind doesn't make much difference. But an ocean is made of
drops, so if you change a lot of them, you change the way the whole
thing works.

Alpha's neurons do a pretty good job of guessing what is a good move,
but because you can't really know what's truly good or bad until you
try it out, Alpha has another trick up her sleeve: she guesses what
the future will bring by rolling dice.

Yes, you heard it right - she just guesses! She follows her guesses
all the way to the end of the game, sees who won, and then sends that
information back down the line (actually, back down the tree of lines,
which branches at every move) to improve her estimates of the values
of her initial guess, and carries on doing that until either she's
pretty sure she's found as good a move as she can, or her automatic
alarm clock tells her to just choose because the clock is ticking.

Lee Sedol and the rest of us should scratch our heads as well as shake
them in dismay, for we all have been beaten at our own game by a
dumbass box of tricks that just guesses. Fancy that!

_________________
i sweat, therefore i stink


This post by djhbrown was liked by 4 people: Bantari, daal, DiceSpin, Pippen
Top
 Profile  
 
Offline
 Post subject: Re: AG Explanation for Dummies?
Post #3 Posted: Tue Feb 21, 2017 11:58 am 
Gosei
User avatar

Posts: 1901
Location: Groningen, NL
Liked others: 189
Was liked: 1017
Rank: Dutch 4D
GD Posts: 645
Universal go server handle: herminator
djhbrown wrote:
This message is written to be intelligible to young children about how
Alphago's mind works.


I'm afraid you've, erm, failed rather spectacularly at achieving that. That mostly wasn't even intelligible to adults. :roll:

Top
 Profile  
 
Offline
 Post subject: Re: AG Explanation for Dummies?
Post #4 Posted: Tue Feb 21, 2017 1:57 pm 
Dies in gote
User avatar

Posts: 52
Location: Germany
Liked others: 0
Was liked: 11
Rank: 4 kyu
Let's see if I can simplify djhbrown's explanation (which is actually not that bad):

djhbrown wrote:
[...] First, let's see how your own mind works, so we can compare the two. [...] Your brain is a [network of] neurons, each of which can [receive input from certain other neurons] and [send a signal] of its own [to certain other neurons]. [...]

The [signals] that neurons [send] to each other are [...] like [...] simple yells: the louder one neuron yells to another, the more likely the other will hear it. [...] Each neuron receives [signals] from hundreds, sometimes thousands, of other [neurons] at a time. [...]

All a neuron [] has to do is [...] do a simple sum of all the 'on' [signals it receives] minus all the 'off' [signals] and see whether that sum is big enough [...]. If it [is], it [sends this information to the other neurons its connected with.]

A neuron's [signal is] either on, or it's off. It's binary, just like the signals flowing around inside a digital computer. Neurobiologists call the [signal] an "axon spike all or nothing response" because they noticed that its intensity and frequency does not change from one [neuron] to another - it's either there, or it's not there.. [...]

And so, finally, we come to Alphago. [...] Alpha has lots of neurons too; they're a bit different from biological neurons but not all that different, in terms of the function they perform. They have their own electronic equivalent of [] receiving and sending [signals] [...]. [Its] function was inspired by the function of the biological neurons [...]. But that's where the similarity ends, for whereas biological neurons can grow and shrink, and live and die, [...] artificial neurons [do not].

Nevertheless, they can be told to learn, to change their [] decisonmaking behaviours. If all goes well, [the neurons] are given a bonus [(they may yell louder)], but if things go badly, they are turned down [(are told to yell quieter)] [...]. One drop in the ocean doesn't make much difference, and changing one [neuron] doesn't make much difference. [But] if you change a lot of them, you change the way the whole thing works.


Some neurons are input neurons, meaning they receive their input from the outside world. In this case it would be the arrangement of the Go board. These input neurons are connected with the neuron network and, eventually the signal will be relayed to certain output neurons, which represent probability values for winning at each possible legal move (i.e. the output of the network is as large as the board).

Quote:
Alpha's neurons do a pretty good job of guessing what [] a good move is, but because you can't really know what's truly good or bad until you try it out, Alpha has another trick up her sleeve: [...] She follows her guesses all the way to the end of the game, sees who won, and then sends that information back down the line (actually, back down the tree of lines, which branches at every move) to improve her estimates of the values of her initial guess, and carries on doing that until either she's pretty sure she's found as good a move as she can, or [runs out of time].

[AlphaGo is incredible!]


This post by yishn was liked by: DiceSpin
Top
 Profile  
 
Offline
 Post subject: Re: AG Explanation for Dummies?
Post #5 Posted: Wed Feb 22, 2017 7:35 am 
Lives in sente

Posts: 853
Liked others: 0
Was liked: 138
I also agree, not that bad, but I do want to make one sort of correction.

Can't make new friends --- that is pretty much true. Which cells are connected to which other cells (can be connected) is defined by the program that is emulating the neral net. In other words, the maximimum connectivity is defined.

But can't lose friends is not ---- think of each connection to neighboring cells having a coefficient of connectivity. It is precisely by changing these values that the neural net learns the function. There is usually no restriction that this coefficient can't go to zero, in effect cutting that interaction off. The connection still POTENTIALLY exists (could be restored) but not at the moment.

Top
 Profile  
 
Offline
 Post subject: Re: AG Explanation for Dummies?
Post #6 Posted: Wed Feb 22, 2017 6:59 pm 
Lives with ko
User avatar

Posts: 214
Liked others: 8
Was liked: 27
Rank: NR
regarding the unintelligibility or otherwise of my message, i would be interested to hear from anyone (or their children) who are not AI cognoscenti, whether your/their imaginations are tickled or confused by it, and would be willing to try to answer in plain English any questions or objections that it provokes.

and those of cognoscenti too, to the extent that i can, eg:
Mike Novack wrote:
can't lose friends is not [pretty much true]

i would say that a coefficient going to zero is more like having a neighbour whose opinions you ignore.

one thing i wanted to mention in my message but didn't, is that embryonic biological neurons move by chemotaxis (and axons grow in directions guided by chemotaxis http://www.mitpressjournals.org/doi/ful ... CO_a_00426), the same way that bacteria navigate, which tickled my imagination when i first heard about it.

it maybe won't be all that long before someone figures out how to endow virtual neurons with this functionality and science takes another step towards smart cyborgs.

Top
 Profile  
 
Display posts from previous:  Sort by  
Post new topic Reply to topic  [ 6 posts ] 

All times are UTC - 8 hours [ DST ]


Who is online

Users browsing this forum: No registered users and 5 guests


You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum
You cannot post attachments in this forum

Search for:
Jump to:  
Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group