Even before AG, a certain Chinese top pro already studied it, and went, "What's so good about this move?", but lacked a number to back up the assessment. Now we have it, a 5% drop in winrate. Off the pedestal.

Maybe it's like tourists going to famous restaurants just because they are mentioned in guidebooks.

There are dozens of other good restaurants around, but the tourists don't know about them. If there is little information about the other places then you'll see people raving about the ones they do know. They'll queue for hours - unlike the locals, which go to other places -, which increases the tourist restaurants' reputations even further - "if it's so good that others queue for hours, I want to go there as well"!

Like much to do with Shusaku, there's a big dollop of hype. The original story had nothing to do with pro merits of the actual move. It was do with the unproven assumption of a doctor who claimed to have seen Genan's ears redden. This reddening may or may not have happened; it may or may not have had anything to do with consternation over Shusaku's move - it may or may not been due to over-imbibing of sake, or sitting next to the brazier, or having lascivious thoughts about the maids. But then the doctor wouldn't have sounded so impressive, would he?

Similarly with the famous dame no myoshu move. Pros have been split over whether it was an acceptable move or an awful waste of a move. Lizzie says it was awful (reverses the winrate 15%), but was Chitoku playing go or was he playing mind games - was he so far ahead (as Lizzie agrees) that he was really just saying to Genjo: "Hey, pal, shouldn't you be resigning? We are missing good drinking time when we could be reddening our ears!"

Lee Sedol's move against AlphaGo in game 4 wasn't probably the best move neither.

Regarding the ear reddening move, is it at least the best move in that part of the board?

In my books, an outstanding move does not need to be the best move. If it turns the game around, it did the job wonderfully. And if some ear gets red, even better. Lee Sedol's move against AlphaGo in game 4 wasn't probably the best move neither.

And if some ear gets red, even better.

But aren't all those old games without komi? If so, asking for those bots, trained for 7.5pt komi, to evaluate board positions or moves, does not seems like a good idea to me.

Could someone usefully comment on what I regard as slightly unusual behaviour by Lizzie/LZ (I am using rounded figures for simplicity):

1. White starts with a projected win ratio of 54%.

2. If Black passes for Move 1, the White win ratio goes up to 77%

3. If Black plays on the 19-19 point on Move 1, the White win ratio goes up to 84%. (And Black moves on the edge give around 76%.)

That seems to say, passing is worth more to Black than playing certain moves on the board. Counter-intuitive?

Abstract

Recent research has revealed that the output of Deep Neural Networks (DNN) can be easily altered by adding relatively small perturbations to the input vector. In this paper, we analyze an attack in an extremely limited scenario where only one pixel can be modified. For that we propose a novel method for generating one-pixel adversarial perturbations based on differential evolution. It requires less adversarial information and can fool more types of networks. The results show that 68.36% of the natural images in CIFAR-10 test dataset and 41.22% of the ImageNet (ILSVRC 2012) validation images can be perturbed to at least one target class by modifying just one pixel with 73.22% and 5.52% confidence on average. Thus, the proposed attack explores a different take on adversarial machine learning in an extreme limited scenario, showing that current DNNs are also vulnerable to such low dimension attacks.