AlphaGo obviously made mistakes in game 4 under the pressure from LSD's brilliant play. I'd like to know if the "dumb moves" are caused by the lack of pro data or some more fundamental flaws with the algorithm/methodology. AlphaGo was trained on millions of amateur games, but if Google/Deepmind builds a website where people (including prop players) can play with AlphaGo, it would be interesting to see who improves faster.
My guess is that Sedol won because he introduced sufficient complexity through cutting points and numerous black groups (see the image). Since AlphaGo uses Value and Policy networks to determine the hot spots to analyse using Monte Carlo tree searches, by making a game rife with lots of simultaneous fights, Sedol dodged the one-two punch of Value and Policy networks combined with MCTS.
In other words, if Sedol can make over a dozen points of interest on the board, AlphaGo cannot deeply assess them all. In the image, there are at least 13 interesting moves and cuts plus up to 15 groups (depending if lone stones are considered groups by AlphaGo). I suspect that this position was far more complex than at any point during any of the three previous games.
It might also explain the meltdown of playing out an unfavourable ladder (the P10 group, as P8 is another possible move).
Eventually, math wins. There will come a point where humans cannot make the game sufficiently complex to beat
a domain-specific machine intelligence (such as AlphaGo).
This seems to be more impressive than the Deep Blue moment. In 1996, Deep Blue didn't make it on the first try. Even in 1997, it has been a draw until the 6th game. Although AlphaGo has 2 games to go, the first three seem to be a clear victory.
This time the computer did not win out of pure bruteforce. Deep Blue relied on an opening book and massive computational power to explore the game tree. After the opening it was pretty much on its own, bruteforcing moves.
This technology used a neural network trained with hundreds of thousands of games which provided the pattern matching aspect, combined with the bruteforce move sequence reading, the montecarlo tree search... and 1200 CPUs + 600 GPUs.
Assuming a Titan X with single precision, those 600 GPUs are 4 PFlops! Deep Blue extrapolated to today with Moore's law would only be ~72TFlops.
While DNN+RL+Tree search is cool, the hardware requirements for AlphaGo to play at this level are staggering and only supported by large marketing budgets :)
AlphaGo is using a 1920 cpu/ 280 gpu distributed setup for the Sedol games. One source reported the gpus are Nvidia K40. The gpus give a peak possible performance of 470 terraflops. That would put it somewhere in the middle of the Top500 list, similar to Deep Blue in its time.
Note though that AlphaGo almost certainly uses single precision arithmetic -- for neural networks even single precision is overkill.
Also the Top500 list is based on Linpack, which measures performance for computations that are pretty strongly interconnected across the different processors of the system. AlphaGo's Monte Carlo tree search problem is more embarrasingly parallel, with evaluation of different positions really being independent computations.
It is much easier to make systems that can handle embarrasingly parallel loads than the highly interconnected loads handled by the top500 supercomputers. So even though the flops are comparable, the systems are not.
