> comparing AlphaGo Distributed to, e.g. CrazyStone, is to compare two distinct states of the art – performance given small computational power (and a small team, for that matter) and performance given massive computational power and the efforts of over a dozen of the best AI researchers in the world.
I know the author probably doesn't have any stake on it, but defending that the state of the art should be categorized by the size of the team of researchers behind each product sounds like just being butt-hurt. "Look what this team accomplished with only two people" only compares the researchers, not the end results.
I totally agree with you with respect to the size of the team, but I do think that the computational power is a very important factor. This is the cost factor, both for investment in hardware, power bill and maintenance of the "cluster". If you have a very large cluster, you do not have one program who won, but one program plus a team ensuring that the cluster is running and healthy against a single go player.
I didn't read it that way. That authors' goal is towards assessing progress in the field. To some extent, adding more researchers is like adding more CPUs; fuel for an industrial process, which should be corrected for when trying to gauge the rate of return on those investments at the current level of technology.
>Hassabis at AAAI indicated DeepMind’s intent to try to train AlphaGo entirely with self-play. This would be more impressive, but until that happens, we may not know how much of AlphaGo’s performance depended on the availability of this dataset, which DeepMind gathered on its own from the KGS servers.
As an amateur (10k) go player who watch the Fan Hui games, I can say that AlphaGo seems to rely heavily on the training data. This is because all of its moves feel very human, even in circumstances where strong humans find better, weird looking moves. This is in contrast to chess AI (and even other go AI) that feel distinctly roboty in how they play. In watching a professional (9p) commentary of the games, it seems that Fan Hui not lose because of particuarly good moves of AlphaGo, but because of some specific mistakes that he made (this is not unusually, as most professional games come down to loosing moves, instead of winning moves). In this sense, AlphaGo seems to be playing at human level, but with fewer mistakes.
This is certainly impressive, but unfortunately AlphaGo currently seems to be a demonstration of synthesizing and automating expert human knowledge, instead of creating new knowledge.
I don't know what his point was that it was predictable. If you were closely following the field, sure. There were some super promising papers published at the end of 2014. Heck if you just saw the wave of deep learning results that were crushing various AI domains, you could have predicted Go would be on the chopping block soon. I was really sure that Go would be beaten by the end of 2015, which didn't happen. However almost everyone I spoke to was skeptical of that, and the outside world was totally unaware. And it still represents relatively fast progress, even if the progress is predictable.
Anyway I think drawing strong conclusions about AI progress from this single example is wrong. You need to look at general trends, which is definitely rapid improvement.
>Chess programs have improved by around fifty Elo points per year over the last four
decades. Estimates for the significance of hardware improvements are very noisy but are
consistent with hardware improvements being responsible for approximately half of all
progress. Progress has been smooth on the scale of years since the 1960s, except for the
past five.
>Go programs have improved about one stone per year for the last three decades. Hardware
doublings produce diminishing Elo gains on a scale consistent with accounting for
around half of all progress.
While I think I actually said similar things about this being incremental progress when AlphaGo came out, what I would really enjoy would be an evaluation of how much progress the enhanced techniques represent as AI techniques. AlphaGo involved a Convolutional Neural Network driving Monte Carlo Tree Search with a linear approximator thing to speed things up. Convnets are themselves something of a hybrid system.
So, will the future look like combining together more and more hybrid levels? Is there a limit to the number of layers or pieces? Maybe only someone Yann Lecun could comment. Still, I'd be curious.
In general, training requires far more neural networks requires far more computational power than using them does. However, AlphaGo combines neural networks with a traditional Monte Carlo search. This means that you can improve the performance of a trained AlphaGo by just giving it more processing power. Indeed, with a sufficient amount of processing power, AlphaGo would converge to fully optimal play.
It's not really the traditional Monte Carlo search; IIRC they're using UCT weighting for MCTS. Apologies if that's what you meant, but I think to most people "traditional Monte Carlo" means something different (probably uniform depth charges).
You're not going to be able to "rigorously model AI progress" at all well if you don't start by quantifying exactly what deep learning techniques are actually doing, and what resources they require to do it.
I know the author probably doesn't have any stake on it, but defending that the state of the art should be categorized by the size of the team of researchers behind each product sounds like just being butt-hurt. "Look what this team accomplished with only two people" only compares the researchers, not the end results.