Life In 19x19 :: DeepZenGo thread

I have seen a lot of Deepzen training games and some of them are really interesting. How about a study group on Deepzen?

Deep neural networks are highly expressive models that have recently achieved state of the art performance on speech and visual recognition tasks. While their expressiveness is the reason they succeed, it also causes them to learn uninterpretable solutions that could have counter-intuitive properties.

Second, we find that deep neural networks learn input-output mappings that are fairly discontinuous to a significant extent. We can cause the network to misclassify an image by applying a certain imperceptible perturbation, which is found by maximizing the network's prediction error. In addition, the specific nature of these perturbations is not a random artifact of learning: the same perturbation can cause a different network, that was trained on a different subset of the dataset, to misclassify the same input.

Uninterpretable solutions with counter-intuitive properties? AlphaGo, anyone?

There is a good example, for visual learning, of such perturbations in this Quanta Magazine article, "Is AlphaGo really such a big deal?" ( https://www.quantamagazine.org/is-alpha ... -20160329/ ). Look for the picture of a dog. Perhaps, in go, the go position where DeepZen made a kyu level life and death error is one such perturbed position. My guess is that Deep Zen would have no trouble with the life and death problem by itself, or in other whole board contexts, but in this case the top side introduces a discontinuity so that the bottom side group got ignored.

I feel if it would be this easy for a shape to mislead the NN, then bots would still be much weaker.

1st row says Shibano Toramaru 7 dan 7 wins.

I think on the left side shows pro who win the most against it, sadly I can't read who is who
The right side seems to be accumulate results, 1251 wins and 68 loses, 95% winrate

I saw this shape a lot and deepzen is always playing the hane.

Zen made what looks like a pretty basic life and death mistake in the game it's currently playing against Sada Atsushi

Author:	Uberdude [ Mon Oct 02, 2017 10:04 am ]
Post subject:	DeepZenGo thread
From this thread on a cool move the bot DeepZenGo played: lichigo wrote: I have seen a lot of Deepzen training games and some of them are really interesting. How about a study group on Deepzen? Volia! Team human got a rare win today with a kill that Zen missed. I wonder if it totally did not consider this false eye making tesuji (as it's so rare to actually happen and work it could have low probability in policy network), or did it misread the continuation? Black's previous move after the fighting on lower side was the push on the top side (g15). Attachment: killZen.png [ 182.45 KiB \| Viewed 16658 times ]

Author:	jeromie [ Tue Oct 03, 2017 8:08 am ]
Post subject:	Re: DeepZenGo thread
Thanks for the thread! I love it when a textbook tesuji like that works; it feels like all those hours of tsumego really paid off. That's kind of a surprising mistake to me. I'd think a big cut like that would show up in the policy network, and the reading doesn't look too difficult. Given how MCTS works, though, I could see the kill being hidden beyond the horizon of its search.

Author:	pookpooi [ Tue Oct 03, 2017 11:43 am ]
Post subject:	Re: DeepZenGo thread
Very good news http://www.usgo.org/news/2017/10/ai-goo ... day-night/

Author:	Bill Spight [ Thu Oct 05, 2017 7:45 am ]
Post subject:	Re: DeepZenGo thread
I recently ran across a couple of articles on deep neural networks which may be pertinent. One is called "intriguing properties of neural networks" ( https://arxiv.org/abs/1312.6199 ). To quote from its abstract: Quote: Deep neural networks are highly expressive models that have recently achieved state of the art performance on speech and visual recognition tasks. While their expressiveness is the reason they succeed, it also causes them to learn uninterpretable solutions that could have counter-intuitive properties. Uninterpretable solutions with counter-intuitive properties? AlphaGo, anyone? Quote: Second, we find that deep neural networks learn input-output mappings that are fairly discontinuous to a significant extent. We can cause the network to misclassify an image by applying a certain imperceptible perturbation, which is found by maximizing the network's prediction error. In addition, the specific nature of these perturbations is not a random artifact of learning: the same perturbation can cause a different network, that was trained on a different subset of the dataset, to misclassify the same input. There is a good example, for visual learning, of such perturbations in this Quanta Magazine article, "Is AlphaGo really such a big deal?" ( https://www.quantamagazine.org/is-alpha ... -20160329/ ). Look for the picture of a dog. Perhaps, in go, the go position where DeepZen made a kyu level life and death error is one such perturbed position. My guess is that Deep Zen would have no trouble with the life and death problem by itself, or in other whole board contexts, but in this case the top side introduces a discontinuity so that the bottom side group got ignored.

Author:	moha [ Thu Oct 05, 2017 8:49 am ]
Post subject:	Re: DeepZenGo thread
Bill Spight wrote: Uninterpretable solutions with counter-intuitive properties? AlphaGo, anyone? While AG did come up with surprises, it's harder to recall examples where it's choice cannot be explained by deep exploration of variations. I have seen many such reviews, and the surprise factor is usually where it can ONLY be explained like that, so where the strategic meaning is missing ("such things should never work - we thought"). Quote: There is a good example, for visual learning, of such perturbations in this Quanta Magazine article, "Is AlphaGo really such a big deal?" ( https://www.quantamagazine.org/is-alpha ... -20160329/ ). Look for the picture of a dog. Perhaps, in go, the go position where DeepZen made a kyu level life and death error is one such perturbed position. My guess is that Deep Zen would have no trouble with the life and death problem by itself, or in other whole board contexts, but in this case the top side introduces a discontinuity so that the bottom side group got ignored. I feel if it would be this easy for a shape to mislead the NN, then bots would still be much weaker. But even without special perturbations, no forward pruning can be 100% correct, so rare oversights will always happen. Maybe they don't normally show up directly like this (around top of tree they can be corrected by good search techniques), just weaken the bot by damaging the evaluation of deeper nodes, unnoticed. This may be something where AG differs from other bots: I sometimes feel many others just take the top 3-4 picks from the NN and doesn't search other moves at all, even at the top, which would explain the number of tactical artifacts popping up in their games.

Life In 19x19 http://lifein19x19.com/

DeepZenGo thread http://lifein19x19.com/viewtopic.php?f=15&t=14576	Page 1 of 4

Author:	Bill Spight [ Thu Oct 05, 2017 9:04 am ]
Post subject:	Re: DeepZenGo thread
moha wrote: I feel if it would be this easy for a shape to mislead the NN, then bots would still be much weaker. Who said it was easy?

Author:	Uberdude [ Fri Oct 06, 2017 6:50 am ]
Post subject:	Re: DeepZenGo thread
Just a general comment on something I found interesting with these Zen games. Many humans have been trying the AlphaGo-style early 3-3 invasions under 4-4, but Zen doesn't play them itself, so has obviously not reached the liking of them AlphaGo has. However, when played against it, Zen will often play the one point jump instead of hane after the opponent crawls as Alpha does, and it often ends up with one of the AlphaGo 3-3 invasion joseki. I don't know if Zen has been trained on AlphaGo self-play (very small sample, unlikely to have much effect I think) or human games featuring the AlphaGo-style 3-3, but it's interesting that it has (at least fairly independently) reached the same conclusion that when the opponent plays early 3-3 jump instead of hane is often a good reply.

Author:	pookpooi [ Mon Oct 09, 2017 1:37 am ]
Post subject:	Re: DeepZenGo thread
DeepZenGo record at last week I think on the left side shows pro who win the most against it, sadly I can't read who is who The right side seems to be accumulate results, 1251 wins and 68 loses, 95% winrate

Author:	Uberdude [ Mon Oct 09, 2017 1:50 am ]
Post subject:	Re: DeepZenGo thread
1st row says Shibano Toramaru 7 dan 7 wins.

Author:	lichigo [ Mon Oct 09, 2017 5:38 am ]
Post subject:	Re: DeepZenGo thread
Everyday I watch it and Deepzen's moves are surprising ^^ Would be a good subject for invisible 2 hhahah

Author:	John Fairbairn [ Mon Oct 09, 2017 5:52 am ]
Post subject:	Re: DeepZenGo thread
Quote: 1st row says Shibano Toramaru 7 dan 7 wins. And his brother has 3 wins. It would be interesting to know whether they've conferred and found a weakness.

Author:	Uberdude [ Mon Oct 09, 2017 6:41 am ]
Post subject:	Re: DeepZenGo thread
Shibano is getting smashed by Zen as I write this... All his big (almost) territories grovelled down to 2 eyes whilst Zen's remain intact. Its feel for the topology of centre fighting is top notch.

Author:	Bill Spight [ Mon Oct 09, 2017 7:34 am ]
Post subject:	Re: DeepZenGo thread
BTW, from what I hear, human-computer teams in chess are better than computers alone, despite the fact that the computer programs are very much better than humans. I don't mean like pair go, where the program and human alternate plays; I mean where the human makes the final choice of each move. That may have something to do with the different basis of the chess programs, by comparison with go programs. IIUC, the chess programs rely heavily upon search. It would be interesting to see if that is so in go. Such games might be valuable, not just to see if it is possible, but also to discern possible weaknesses in computer play and strengths in human play. It seems likely to me that a pro plus DeepZen would be stronger than DeepZen alone. And maybe even an amateur shodan plus DeepZen would be stronger. It could also be good training for the human, to learn to tell which of DeepZen's suggestions is best. It might even be better for training amateurs than pros.

Author:	bernds [ Tue Oct 10, 2017 8:07 am ]
Post subject:	Re: DeepZenGo thread
pookpooi wrote: I think on the left side shows pro who win the most against it, sadly I can't read who is who The right side seems to be accumulate results, 1251 wins and 68 loses, 95% winrate Is there an archive of these games somewhere?

Page 1 of 4	All times are UTC - 8 hours [ DST ]
Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group http://www.phpbb.com/

Author:	Uberdude [ Tue Oct 10, 2017 8:13 am ]
Post subject:	Re: DeepZenGo thread
Go4go has about 150 of Zen's games (mostly from this event): http://www.go4go.net/go/games/byplayer/1776

Author:	lichigo [ Sat Oct 21, 2017 9:00 pm ]
Post subject:	Re: DeepZenGo thread
Deepzen is on tygem, it is not new but the 20th october tygem announced that deepzen will play for a year and available 24/7. Since the 20th deepzen played (at the moment) 56 games and lost only 2 (22 and 35 wins in a row). Far to win 60 games in a row or the perfect like Master I think deepzen is an excellent sparring partner. Some players tried to play some alphago zero moves or openings. Deepzen won agaisnt a lot of strong pros included Park Jeongwhan and some young chinese pros. Hope we can share some moves and ideas. I think we can also learn from Deepzen ( sometimes it is copying some of the Alphago moves).

Author:	Uberdude [ Sun Oct 22, 2017 2:32 am ]
Post subject:	Re: DeepZenGo thread
Zen made what looks like a pretty basic life and death mistake in the game it's currently playing against Sada Atsushi and I think it might lose given some typical Monte carlo tilt suicide moves. The hane below should be placement one to the right to avoid black's throw in and ko for life. I keenly remember this shape as I made the same mistake in one of my online BIBA league games (and indeed I lost the game from this ko aji). It's not a totally unconditional mistake in that the hane does have some advantages, such as black being less likely to peep the outside cutting point of the wall because white could play atari at b3 and that destroys the ko whilst building eyeshape, and the j3 cut means the group isn't totally dead yet, but isn't this a big mistake seeing as black got the 3 marked tenukis out of the sacrifice so leaving a ko seems bad? Click Here To Show Diagram Code [go]$$Wc DeepZenGo (white) vs Sada Atsushi 3p $$ +---------------------------------------+ $$ \| . . . . . . . . . . . . . . . . . . . \| $$ \| . . . # . . . . . . . . . . . . . . . \| $$ \| . . . . . X . . . . . . . X . . . . . \| $$ \| . . . O . . . . . , . . . . . , X . . \| $$ \| . . . . . # . . . . . . . . . . . . . \| $$ \| . . O . . . . . . . . . . . . . . . . \| $$ \| . . . S . . . . . . . . . . . . . . . \| $$ \| . . . . . . . . . . . . . . . . . . . \| $$ \| . . . . . . . . . . . . . . . . . . . \| $$ \| . . . , . . . . . , . . . . . , O . . \| $$ \| . . . . . . . . . . . . . . . . . . . \| $$ \| . . . . . . . . . . . . . . . . . . . \| $$ \| . . . . . . . . . . . . . . . . . . . \| $$ \| . . . . . . . . . . . . . . . . . . . \| $$ \| . . . . . . . . . . . . . . . . . . . \| $$ \| . O O O . . . . . , . . . . . X . . . \| $$ \| . . X X O O O O . . . . . . . . . . . \| $$ \| . O X . X X X X O . . . . . . . . . . \| $$ \| . . 1 . . . . . . . . . . . . . . . . \| $$ +---------------------------------------+[/go]

Author:	lichigo [ Sun Oct 22, 2017 2:39 am ]
Post subject:	Re: DeepZenGo thread
I saw this shape a lot and deepzen is always playing the hane.

Author:	Uberdude [ Sun Oct 22, 2017 3:00 am ]
Post subject:	Re: DeepZenGo thread
lichigo wrote: I saw this shape a lot and deepzen is always playing the hane. Interesting: I wonder if the humans are playing this sequence on purpose knowing Zen has a habit of making a mistake to get an early advantage. Or maybe this mistake is not as bad as black's invading and then dying that it's still actually ok for white overall. Zen won by 2.5 anyway...

Author:	moha [ Sun Oct 22, 2017 4:50 am ]
Post subject:	Re: DeepZenGo thread
Uberdude wrote: Zen made what looks like a pretty basic life and death mistake in the game it's currently playing against Sada Atsushi This may be related to the other LD mistake discussed earlier here. The error may came from the NN and the placement probably wasn't searched at all, or at least not with sufficient effort to understand the consequences. I wonder how the network ranked that move.