It is currently Thu Mar 23, 2017 3:07 pm

All times are UTC - 8 hours [ DST ]




Post new topic Reply to topic  [ 7 posts ] 
Author Message
Offline
 Post subject: Exploring the Frontiers of Knowledge (Demis Hassabis)
Post #1 Posted: Tue Feb 28, 2017 2:43 am 
Beginner

Posts: 9
Liked others: 0
Was liked: 5
Rank: KGS 3kyu
An amazing talk from Demis Hassabis about AI (and Alphago) and it's future : "Exploring the Frontiers of Knowledge"

https://www.youtube.com/watch?v=Ia3PywENxU8


This post by yharzu was liked by: Bonobo
Top
 Profile  
 
Offline
 Post subject: Re: Exploring the Frontiers of Knowledge (Demis Hassabis)
Post #2 Posted: Mon Mar 20, 2017 4:00 pm 
Oza

Posts: 3839
Location: Cambridge, UK
Liked others: 131
Was liked: 1858
Rank: UK 4 dan
KGS: Uberdude 4d
OGS: Uberdude 7d
This evening I went to a talk in Cambridge by Demis. It was mostly the standard DeepMind story from beginnings through Atari to AlphaGo (the meat of the presentation with an intro to Go) and future plans which is probably familiar to followers of AlphaGo here (a lot of slides in common with the video above, but about 1 hour long). It was recorded (and relayed live to an overspill room, the theatre of 300 was packed) so check the host Cambridge Society for the Application of Research website for that later: http://www.csar.org.uk/. There were however a few new titbits I found particularly interesting:
- After the Lee Sedol match they developed an "anti-AlphaGo" that was designed to pick moves AlphaGo didn't explore (but presumably not crap ones!) to fill in the gaps and avoid over-fitting. [It no longer makes the mistake against Lee's divine move, I don't know if due to this approach]. This adversarial learning is hot in computer vision atm.
- I asked about Master playing a more unconventional opening than AlphaGo, was this simply the evolution of its style or could they tweak parameters. He said there is a parameter they call 'heat' [I've seen another term for this elsewhere] which is how strictly or not it has to pick what it thinks is the best move and they changed this so it might pick the fifth best move if it's only a little worse and as moves in the opening are all pretty similar it did new things. [So is the small shimari very inferior?].
- Master only ran on 1 GPU!
- No comment on does an AlphaGo trained from scratch beat human-game trained AlphaGo, announcements soon.


Last edited by Uberdude on Mon Mar 20, 2017 4:34 pm, edited 1 time in total.

This post by Uberdude was liked by 4 people: Bonobo, Kirby, maceta, Waylon
Top
 Profile  
 
Offline
 Post subject: Re: Exploring the Frontiers of Knowledge (Demis Hassabis)
Post #3 Posted: Mon Mar 20, 2017 4:26 pm 
Lives in gote

Posts: 584
Location: Austin, Texas, USA
Liked others: 37
Was liked: 192
Not much to go on for speculation but that never stopped anyone so...

It sounds similar to what they published in the Nature article where they picked a random position from a game, and then added a single random move. I guess they extended that concept.

An analogy with human study could be during games you mostly rely on your trained instincts (neural nets). But in post game analysis you might spend time exploring a wider range of options. Then you might find some holes in your instincts, and try to patch those up.

Top
 Profile  
 
Offline
 Post subject: Re: Exploring the Frontiers of Knowledge (Demis Hassabis)
Post #4 Posted: Mon Mar 20, 2017 6:00 pm 
Lives in sente

Posts: 817
Liked others: 345
Was liked: 176
Rank: AGA 4k KGS 4k
GD Posts: 61
KGS: dfan
Uberdude wrote:
- I asked about Master playing a more unconventional opening than AlphaGo, was this simply the evolution of its style or could they tweak parameters. He said there is a parameter they call 'heat' [I've seen another term for this elsewhere] which is how strictly or not it has to pick what it thinks is the best move and they changed this so it might pick the fifth best move if it's only a little worse and as moves in the opening are all pretty similar it did new things. [So is the small shimari very inferior?].

The usual term for this parameter is "temperature". Basically, every move that is being seriously considered gets a score s, and the relative chance that it gets played is exp(s/T). For very high values of the temperature T, the exponent will always be small and all moves will have roughly equal probability. For very low values of T, the exponents will be spread out widely, and the best-scoring move will be selected effectively every time.


This post by dfan was liked by: Uberdude
Top
 Profile  
 
Offline
 Post subject: Re: Exploring the Frontiers of Knowledge (Demis Hassabis)
Post #5 Posted: Mon Mar 20, 2017 6:44 pm 
Lives in gote

Posts: 363
Liked others: 25
Was liked: 103
GD Posts: 10
I'm very surprised it ran on only 1GPU
Still, we don't know the number of CPU and TPU yet

Top
 Profile  
 
Offline
 Post subject: Re: Exploring the Frontiers of Knowledge (Demis Hassabis)
Post #6 Posted: Mon Mar 20, 2017 7:36 pm 
Dies with sente

Posts: 74
Liked others: 0
Was liked: 23
Rank: NR
picking the 5th-best move (presumably so long as its heat is still >50%) is a neat zigzag trick to throw off opp's reading of the 4 most sensible lines, throwing a spanner into the works of their logic, causing self-doubt emotional reactions which could negatively impact the coherence of their subsequent reasoning, driving them into a crevasse. Master is indeed well-named - the master of chicanery - although Harlequin might be even more apposite; or even Pied Piper, leading the Herd of Adoring Acolyte blind faithful up a cliff of delusion (as incomprehending apoplectic stone-throwing reactions will demonstrate).

"What's for breakfast?", said Pooh.

_________________
i ink, therefore i spam
"J'ai toujours succombe a la tentation de raisonner pour comprendre" - Peter Hoeg


This post by djhbrown was liked by: Bonobo
Top
 Profile  
 
Offline
 Post subject: Re: Exploring the Frontiers of Knowledge (Demis Hassabis)
Post #7 Posted: Mon Mar 20, 2017 10:18 pm 
Oza

Posts: 3839
Location: Cambridge, UK
Liked others: 131
Was liked: 1858
Rank: UK 4 dan
KGS: Uberdude 4d
OGS: Uberdude 7d
pookpooi wrote:
I'm very surprised it ran on only 1GPU
Still, we don't know the number of CPU and TPU yet

Me too. It was just a quick comment in the informal Q&A at the end but the context was AlphaGo no longer requiring a beast of a computer like vs LSD to play well, so I think the answer is about 1 CPU and 0 TPUs. (Or maybe as he was speaking for a lay audience "one GPU, a big one" meant "one TPU").

Top
 Profile  
 
Display posts from previous:  Sort by  
Post new topic Reply to topic  [ 7 posts ] 

All times are UTC - 8 hours [ DST ]


Who is online

Users browsing this forum: No registered users and 3 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