Exploring the Frontiers of Knowledge (Demis Hassabis)

[yharzu]

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

[Uberdude]

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.

[yoyoma]

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.

[dfan]

- 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.

[pookpooi]

I'm very surprised it ran on only 1GPU
Still, we don't know the number of CPU and TPU yet

[djhbrown]

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.

[Uberdude]

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").

[hydrogenpi7]

I'm very surprised it ran on only 1GPU
Still, we don't know the number of CPU and TPU yet

I concur, color me surprised too. Master deceived us once, I'll believe it when I see it. If no GPU is needed or only one, why not Google open source AlphaGo or at the very least give the public some means to experiment and play with it directly if it really wants to do good for the AI and the GO community like it says?

When everyone on reddit was saying there was zero change Master is AG, I said that my hunch was that Google was not keeping their word when they stated future AlphaGo games would be played under AlphaBot name.

Where is it official (or officially stated) that Master only needs one single GPU or that it only ran on one GPU? Unless it is official from Deepmind/Google, and/or they have proven it to the public via some sort of live demonstration, then I'm of the opinion that I don't think it is possible. I'm open to be proven wrong, but like before, I have a hunch that I'm not going to be turned out wrong about this either.

[djhbrown]

as for number of CPUs, how many does JueYi have?

[Uberdude]

A recording of the presentation by Demis Hassabis I went to (but not the Q&A) is now online at https://www.youtube.com/watch?v=ZyUFy29z3Cw. The future direction and Anti-AlphaGo is from https://youtu.be/ZyUFy29z3Cw?t=37m42s.