A Homage to AlfieThis is a copy of my take on why Alfie is Queen that i made to Spirits of Baduk about a year ago, but even with the subsequent advent of Alfie0, it's not out of date because the new broom sweeps pretty much the same way as the old one, only with tidier bristles now they've got rid of those distracting random rollouts:
(Mike wrote: From my point of view, the wonder is that a neural net with JUST a few thousand mindless "cells" can play go as well as it can).
I wondered about that wonder too, ever since she made her dramatic debut in 2015.
I think i have now figured it out, thanks to the DARPA video. Let's see if i can summarise briefly why Alphago is so good at Go:
The game of Go is played within the confines of an artificial microcosm, namely, a universe of just 361 pixels, each of which can have just one of three values.
There is no dark matter in that universe (other than Swim shadows...), nor indeed any energy at all, because the pixels do not move; they simply appear and disappear, according to the whims of the gods that put them there and take them away.
These gods have their own desires, but for the purposes of understanding why Alpha is the Omega of Go, we can safely ignore them, for they are peripheral to the central issue and thus inconsequential.
The universe of Go is minute, a mere tree of 10up170 microphotographs. Some while ago, Japanese theorists noticed that the universe could go on forever in a kind of Kurt Godel way, with true but unprovable theorems about who won, so they put a stop to all that nonsense by modifying the ko rule so that triple kos were a thing of the past, but not of the future.
The tree of Go isn't a proper tree, but one whose branches can grow into one another, producing a net - but, thanks to those Japanese rulemakers, a net for which the arrow of time is steadfastly linear. Dick Whittington can turn, but Theresa May and Go can't go back (not that either of them wants to).
Now, although the tree is not a tree but a net, for the purposes of mathematical analysis, we can transform the net into a tree without losing logical consistency, consistent with Godel's Theorem.
The root of the tree is a cold, dark, place, a yawning empty void of nothingness, with (like our own real universe) no vacuum energy at all - a contemporary myth which will make quite a few people red-faced in a few hundred years' time when it is proven beyond all reasonable doubt that they are talking bullshit because they live in a world of mathematical mythturbation which has nothing even remotely to do with the way the actual thing works - superposition, my foot!.
The leaves of the tree are the ends of the universe, each of which has just one of two possible states: black or white. My guess is that, the Deep South and Brexiteers notwithstanding, it will someday be proved that black is superior to white. Certainly, in my own case, i know that to be unequivocally true, for black people are stronger than white ones like me, because they can withstand the onslaught of B-frequency solar radiation, whereas i now have to go to the quack every so often to have a piece of arm or hand cut out because it has lost its mind and is growing out of control.
Now, because every leaf has a binary value, if we apply Hawking's reasoning and run the clock backwards, we can figure out by linear extrapolation that the beginning of time - when black makes a Big Bang on the board by planting a stone on it - also has a binary value, for there is nothing white can do to stop the inexorable March of Time flying like a banana to the inevitable End of Days.
Time waits for no man, and there are bicycles to repair and bicycles to ride and golfballs to strike and surfboards to fall off of, and George Galloway GG epistles to enjoy, and Dylan Thomases to read, and floors to sweep, and so many things, so little time, so here endeth the Lesson. Tune in next week, same time, same channel, for the next thrilling episode of Agatha Christie's whodunnit.
PS Just heard - read - that AGF are going to do for KGS what [unfinished]. Now that TM has declared war on Europe, i suppose AGF will be launching Go missiles at EGF (kind of an ambiguous acronym) from their colony across the ditch in a vain attempt to reverse the Battle of Hastings. That's a good thing, because it will keep them busy irradiating the Northern Hemisphere, which is soon to become the Southern Hemisphere when the Poles reverse (any day now) and then NZ will take its rightful place at the top of the map like it always said it was and YUK/US will be downvoted so much that it will dissolve into the Land Beyond the Paile, somewhere near Timbuktu. See you On The Beach.
http://www.pottonandburton.co.nz/store/ ... e-down-maplast time i wrote something for young children about alphadoggie, the first response was from someone complaining that they didnt understand it, and the second from someone else writing cliffnotes, which was rather nice of them, i thought.
however, i dont expect anyone to be fussed to write cliffnotes on #128, since almost by definition, anyone who could wouldnt have read it, as they are too busy publicising their own ideas and/or selves in more profound yellow presses than a mere grafitti wall.
so i will have to do it myself, if i want it to be done, which i do, so here Go-es:
1. the universe of Go is short - just 361 moves deep, give or take a couple
2. it's not that wide either - at most 361
3. it has no cycles
4. it has a binary ending.
these four characteristics are what makes Alphadog possible, for reasons i will explain. but before doing so, i wish to point out that none of them are characteristics of the real world, which has cycles, and no depth or width boundary as far as anyone knows, except the personal boundary of an individual organism like you or me, which does have significant implications for personal decisionmaking, (as i have learned to my own cost), and many many more branches at each tick of the clock than there are elementary particles in any piece of wood and a couple of bowls of stones.
Of the 4 characteristics, the 4th may be in some ways the most significant, because it's this one that makes neural nets effective.
Back in 1969, Minsky and Papert did some sums on the nets of the day, and figured out that there was something no neural net could ever do: see the hole in a doughnut.
How ironic, for being able to distinguish convex from non-convex is step 1 of every baby Go player's lesson ladder. If you can't see eyes, you can't play Go.
And yet, and yet, although all the backpropagation and layers of convolution in the world don't make any difference - and anyone who says otherwise is either a fool or a liar - so Alphadoggie cannot see eyes either,....
...BUT...
....she can tell black from white, because, as DARPA showed us, that's what neural nets do. The simplest of nets - Perceptrons - drive planes between points scattered in Euclidian space. DCNNs do the same, except that they drive planes between manifold transformations of that space. Whether DARPA's "manifolds" are the manifolds of topology i can't say for sure, it's out of my depth, but i think from his pictures it probably is what Launchbury means.
https://en.wikipedia.org/wiki/ManifoldNevertheless (again i'm out of my depth here, it really needs a proper mathematician to check my homework), the one discrimination they cannot make; the thing they still cannot do, not unless every input to every convolution contains the entire image, they cannot tell the difference between a sphere and a torus - they cannot tell whether there is a hole in the doughnut.
Le Cunn was on the right track to use DCNN to tell small symbols like the ones you're reading apart, because you can feed the entire image of an alphabettycal character into every prior of every convolution because they're so small.
But a Go board is a bit bigger than a pixellated character, maybe too big
!!BUT!! that's what makes this all so remarkable - Alfie doesn't need to be able to see eyes, because (a) the universe is finite depth - emphasis on the word "depth" and (b) it has only one of two endings, making bifurcation possible.
Goodness, i thought it would take me several postings to explain it all, but i've done it in one go. It's just like the history of religion - i had thought it would be very intricate with lots of crossovers, but it turned out that i could fit the entire history of all the world's religions into one table on one A4 page, because the genealogy is linear (Jesus is another name for Horus = Apollo = Sol Invictus = Stonehenge = the Siberian Father Christmas = the Mayan sun-god = etc etc, and Mary = Isis = Ishtar = Easter = Inanna = the Moon = daughter of the Moon = Venus = the length of the human female menstrual cycle = where babies come from).
Oh, except for explaining why Ke Jie's neural net can't do it too. I will do that tomorrow.
PS and i need to explain why, even though he can't see as well as Alfie, he can, like you and me, and unlike Alfie who can't, he can see eyes. How he does that, how you do that too, is all to do with frogs...
PPS the A4 page is here:
https://www.youtube.com/watch?v=LdYdEaA ... 0h&index=8maybe xor is solved, but the topological distinction isn't. it isn't. it theoretically cannot be. but i'm not good enough with pencil and paper to draw all the squiggles needed to prove that to the satisfaction of the editor of Ars Mathematica - if you are, i would love to see it written out; it might be easier to do than i think.
Alfabuttery cannot see eyes, period. she cannot tell the difference between "C" and "O". Le Cunn can, because he puts all the corners of the board into the picture - but alfie doesn't.
Ke Jie and you can - but i'm not going to tell you why until tomorrow, because it will take too long and like i said, there are bicycles to feed, and stroke, and i dont want to spend my whole bloody life sitting at a keyboard, expecially as i've now eaten both croissants and am ready to go out and face the day and pay homage to Horus.
But i will say just this: yes, eyes are indeed theorems that can be proven from axioms - as it happens, that's exactly what Swim's colour map does.
But alfabetter doesn't. she can't. BUT she doesn't have to!!!!!!!! Because she can see all the way to the end of time, and tell black from white, and back it up back to the beginning, to know which step to probably take to probably get to where she wants to go. yes, i did miss out all the stuff about Monty-CarloPythin, which i suppose i have to say, but not now, not now.
BTW alfie doesn't bother me - she excites me, she excites my imagination. Or, rather, she did, until the penny dropped, and now that i get it all, she is as dull as dishwater to me and i'm ready to move on to more interesting things, such as putting Kismet's head on top of Nao's body and putting icGo into her mind, to make an emotional Go-playing robot with consciousness and empathy. Kismet is featured in "Common Ground" - my most important movie ever, my legacy to the world for the benefit of future mankind, which only 14 people have ever seen even though it's been there for years.
https://www.youtube.com/watch?v=rIcOpCx ... 10&t=1172sAnd there's still the problem of how she can learn, but by following in the footsteps of Isaac Newton, Niels Bohr, and Albert Einstein and forgetting about it, just letting my mind calm itself under a tree, or at the racecourse, or in a tram in (forgot the name of the town, now, i'm sure you know which one i mean - Basel, was it?, you know, the one with the big clock at the end of the street) my subconscious will do its stuff without my help and the answer will spring out like a rabbit from a hat tomorrow morning when (if) i wake up again.
the reason alfie (and JueYi, which i figure to be a clone of her) is God of Go is
(a) because Go has finite depth and ends in a binary result, CNN is good enough. DCNN even better, because the search is better constrained. RL essential for statistically sensible guided search.
+
(b) Monte sampling is good enough for a homogeneous space (ie populated by objects of a single type - in this case, 19x19 pixels)
(c) the more grunt you have, the more you can search, so the better your chances of not missing out on anything good that happens to be within reach. With oodles of doodles, Google can supply a truckload of truckloads of processors to keep the peasants at bay, as can anyone who can dole out 5x$300k from their spare change for a week's holiday without blinking an eye.
(d) Poor old Ke Jie can't compete with that, because for all his brilliance against other H.Saps, he can't read ahead 361 moves, and alfie, although she can't read it all, can read wide enough to read him out of his depth, and her DCNN can eliminate a lot of the nonsense width that no sane bot not losing would contemplate, so her chances of lucking in on a good enough defence to anything he might try are better than his chances of lucking in on one of her weak spots, as Lee Sedol managed to last year, to come back like a zombie from the cemetery.
This, despite the fact that alfie can't see eyes and KJ can. He can, because his eyes work like a frog's, because both his and a frog's eyes are built to the same basic formula.
That formula includes two basic things:
1. a fovea (actually, 2 foveas, one each side of the nose)
2. muscles to move the focus of the fovea around the scene by squishing the lens and rolling the balls
KJ and the frog use their motion detectors to notice changes as their focus shifts, and thereby to detect edges, and thereby to detect eyes.
Swim sees eyes in a more plodding sort of way, by growing a colour-connection map iteratively (the equivalent of a recursive definition of a connected cluster working backwards), and then walking around inside it, noting the internal colour-controlled spots as it goes.
The reason i mention frogs is because Hubel and Wiesel were the first to discover that some neurons attached to the retina of a frog fire only when a line of them notice a change from light to dark (or vice-versa) at the same time, thus telling the frog's brain that an edge is moving in a certain direction.
https://www.researchgate.net/publicatio ... hat_the_Fr…
They were careful to say that they weren't saying that frog's eyes and human eyes were the same, as it was in those days still unfashionable to cast humans in the same basket as other beasts of the field - and besides, they hadn't dissected any living human eyes (or at least, they didn't say they had... you never know with mad scientists what they do behind closed doors, but no matter how Frankenstinian they are, they're a whole lot less harmless than Bossy Bankers).
Thou shalt have no other God than She who must solemnly be Obeyed, as Rumpole and Basil put it, so it's no wonder there are so many ardent disciples ready to stone anyone that dares take Her name in vain, as recorded in the Life of Brain, and nightly on the Googlebox.
Biographical Note
in 1976, i entered a new world, the world of the Lecturer, a step up
from Assistant Lecturer (another name for donkey worker) and looked forward to a new life free from the tedious moronities of programming, which i could delegate to lesser peasants (which in Newcastle in 1988 i actually managed to do, lucking in on two such ******** - Bernie and nBertrand - who spoke geekspeak well enough to do in 5 minutes what would have taken me 50 years)
Here's an example of the sort of **** programmers have to put up with (Tiger used to say that engineers are a breed apart, which is so true, but they are closer to normal humans than are software engineers, who are ******* ******).
It occurred when attempting to escape from the clutches of Fedora into the greener arms of KDE Neon, but unfortunately Fedora's open source ***** have ****** up its configuration so much that it's as useless as Xubuntu - 10 years ago, both of them used to work, but then the geeks moved in and ********** it all up.
All because kdenlive, which used to work, no longer does because some **** thought they knew better, which they didnt.
Code:
[d@localhost ~]$ su
Password:
[root@localhost d]# ./home/d/Downloads/RosaImageWriter/RosaImageWriter
bash: ./home/d/Downloads/RosaImageWriter/RosaImageWriter: No such file or directory
[root@localhost d]# /home/d/Downloads/RosaImageWriter/RosaImageWriter
No protocol specified
QXcbConnection: Could not connect to display :0
Aborted (core dumped)
[root@localhost d]# **** you
bash: ****: command not found...
Install package 'the****' to provide command '****'? [N/y]
Tempted though i was, i didn't.
Update 2018 (Good Grief, is it 18 already?! - where on earth did all the time go??): Between then and now, Fedora 27 came out and now kdenlive works again. it still messes up some things, like changing the shape of a frame when you take a snapshot and distorting the image, but luckily VLC doesn't, so there is a workaround.
Writing code that works is really hard, so i guess it's unfair of me to complain about others' hard work so much, especially as i've never done a flaw-free honest day's work in my whole life and now don't do any work at all, except on the pendulum practice range.