I think this is a common misinterpretation of AI bots. They have no concept of initiative or any other human concept of Go. They are merely playing to get one more point than their opponent and are evaluating the position solely in these numeric terms.

Sadly, this also means that they cannot explain why they make a move in a way that we can understand and use.

Perhaps all we can do is the same as Mark: play through (AI) games without attempting to 'understand' what's going on and just trying to feel the flow?

Thanks for the salutary reminder, Jon. I think we do know this* but the urge to anthropomorphise overtakes us and we do end up forgetting. Does my dog really love me or has he just worked out how to get food and shelter? We prefer to believe the former.

But does it not follow from your, dare I say it, cynical view that even trying to follow the AI flow is also bound to fail? What I am getting at is that there may be no flow to follow. The bots' evaluations are always in the moment. There is no detectable (or consistent) strategy behind them. That leaves us humans like fish out of water. Getting back into the pool with humans pros may be the best option, no? Just use AI for ideas and motivation, not as gospel.

*With the caveat that there may a version of katago that tries to optimise the points difference????

PS I've just met the second person in my life I know who plays croquet!

The universe doesn't know it follows Newton's, pardon, Einstein's laws of mechanics, or Maxwell's equations in the electromagnetic field, or Schrodinger's or Dirac's equations. The universe doesn't know the speed of light, Planck's constant etc etc. The universe simply IS.

All of these are human concepts which help us understand the physical and cosmological reality. The fact the universe is unaware of the concepts according to which it works should not discourage us investigating and conceptualizing.

The same happens when we try to understand Go. AI acts as a magnifying glass, or an idiot savant, or whichever analogy you like, for us to understand Go better. We do so collectively and need language, hence terms, to discuss and articulate.

So when we say "AI favor sente" we don't really think AI consciously tries grabbing sente, althought I wouldn't be so sure about what's really in that neural network when left to its own resources and what isn't.

Now back to John, who says AI wants "control" rather than the next move and finds Western go players think of having the next move more than Eastern (pro) players (or AI) and this may be induced by our chess intuition.

I can't speak for the rest of you but I doubt my brief and youthly exposure to Chess still dominates those 25+ years of intensive Go playing and studying. When I talk about sente (maybe wrongly so, but I try to stick to the more common terms, even if they are Japanese) I really mean "the right (and duty) to play elsewhere because the local situation is not so big/urgent anymore (i.e. temperature has dropped)". That is a sharper notion than mere "control". Of course you want control. It probably means you are leading. But the obnoxious striving for the right to play elsewhere is what has been reinforced and encouraged by studying AI and what I find a useful sharpening of my approach.

My emphasis above. I agree with the approach to using AI just as a pointer to where we might have made a mistake and, in addition, where a better move might be. There is no "flow" in AI play because Ai has no grasp of strategy on any large scale. I don't know much (anything?) about how AI programs choose moves but if there is any strategy it must lie in the heuristics (algorithm?) the AI uses to choose what moves to try for playouts and how many playouts to use. We use strategy as a compass in determining where to play and our feeble ability to playout (read) is how we decide which particular move to make. Our pathetic ability to read requires that we have higher level concepts and strategy to compensate.

If you're in a room with a human pro go player, and you don't share a common language to communicate, is it all that different than reviewing with computer AI?

In both cases, the only explanation you can get are the sequences they recommend. No explanation can be given since there is no shared language but the moves themselves.

Still useful, in my opinion.

_________________
be immersed

I wouldn't make such strong claims about what concepts are and aren't represented in the neural nets that drive modern AI.

I once trained a net purely to predict the location of the next move - the net was given literally no data about the win/loss outcome of the game or the score or even that Go was a 2-player game with some objective. Just: see a pattern of stones, identify a location on the board. See another pattern, identify a location.

Later when inspecting the net, I found within it a channel that computed who was ahead or behind in the game and whether by a lot or a little. Noisily and imperfectly, but definitely that's what it was doing - visualizing it in many board positions you could see it consistently always assigning positive weights on territory of one player, negative weights on the other (including correct handling of life and death for dead groups!), just prior to a layer that would sum/average these values across the board into one value before feeding into further processing later.

So without being told anything about winning/losing or the score, purely in service of learning this complex and opaque mapping from "pattern of stones" to "location on the board", the neural net on its own internally discovered and learned to compute a concept that directly corresponds to what humans would name "being ahead or behind". Presumably this is worth computing because humans on average play differently when "ahead" versus "behind" so computing it helps you make better predictions, but the net wasn't told such a concept existed - it came up with it on its own!

There is no doubt that encoded within a strong net's weights and activations, whether trained on human data or trained purely in RL, are many other strategic concepts being computed and represented in different ways, many of them close enough to human concepts to be recognizable if only we had better ways of finding and extracting them. It is just that our current tools are not up to the task (and for the same reason that our tools are not up to the task of deciphering the electrical impulses in a human brain in all but the crudest ways - the task is hard!)

It's also a bit questionable to think of deep-learning-based AI agents as evaluating things "solely numerically". This is probably, as best I can phrase it, confusing the "software" of the weights and activation patterns that encode how a neural net is responding to what it observes with the "hardware" of the fact that we just so happen to be representing these values numerically and crunching them via heavy GPU arithmetic. The "hardware" isn't really fundamental to what the "software" is doing at a conceptual level and will likely lead you to the wrong intuitions about where the strengths and weaknesses of neural nets and deep learning are.

A better way to think of deep-learning-based AI agents is that when properly trained, they tend to evaluate things in a fuzzy and *analog* (i.e. non-digital, non-numeric) way, based on variously learning complex pattern recognition, high-level holistic features, or other concepts depending on the task, at the same sort of level that human instinct and snap judgment is capable of learning. With of course many non-human-like quirks (most prominent when dealing with corrupt or out-of-distribution or certain kinds of "artificial" data), but still, this is the most useful baseline intuition to start with.



Well with a human pro, even with no shared verbal language, if the pro starts playing out something that looks like a tewari, or adjusts some stone positions to show how the same move works (or doesn't) in some alternate situations, I'd expect to sometimes be able to follow along. I don't know of any Go AI that currently does such things to explain itself. (wink)

True. This is a non-verbal capability humans have, which I hadn't thought of. Maybe a potential feature for KataGo

Though, I suppose there are some challenges, such as having the capability to produce alternate board positions that are common and/or relatable to the student.



This write up is fascinating. I feel like I should learn more about neural nets by doing some sort of a simple project.

On one hand, I have been turned off by a lot of ML since it seems like such a black box - I understand what's happening a lot more clearly with traditional "non-ML" programming.

But the power of this stuff is amazing.

_________________
be immersed

Defining activations based on how close the game is necessary to accurately reflect the training data which is also based on how close the game is. And surely there are activation functions that corresponding to every game concept and strategy. But simply extracting these concepts from the NN will at best provide a list of top contributing features. Which is definitely helpful and there is on going work in this area. But this is not enough to providing "teaching" on its own.

Finding concepts recognizable to humans from a board position would require several different ML models specifically training to do so. Models not just for next plays, sequences, and scoring, but for classifying life and death, finding weak points, expressing influence, thickness, moyos and so on. I haven't seen anything like this but it could certainly be done with a lot of effort.

However, a teaching algorithm be needed to somehow present the information to be learned by a player. Simply labeling concepts with a percentage and presenting variations and sequences is not teaching. And teaching must be fit to the student and so models would need to be trained on the player's own games and other player's games by rank.

All this to say, I think we are pretty far away from AI providing actual teaching. Still, there is a lot that could be learned from top contributing features and it would be great that were developed for Go.

I think it's worth reiterating some of lightvector's post:



My takeaway from this is that teachable concepts exist within these neural networks - but they are difficult to expose/illustrate to the observer. In the meantime, we can still benefit from seeing several examples of these concepts.

I wonder if there are ways to make it easy for the end user to visualize/understand the intermediate layers within the neural network. That could potentially give more direct insight into the end percentages and sequences that the network ultimately outputs. IIRC, there was some research done for this in neural networks used to recognize images - intermediate layers of the network gave insight into the overall recognition process.

Maybe more research into exposing these intermediate layers of a network could be beneficial for eventually producing something that exposes the underlying concepts behind a learned algorithm.

_________________
be immersed

I'm not an expert on this topic, so I may be misunderstanding, but I think this is an example of that kind of "intermediate visualization": https://www.mathworks.com/help/deeplear ... twork.html



You can visualize the different activations at various levels. Maybe that kind of thing can give more insight into how the network is learning, overall.

For go, maybe this means more easily visualizing various layers that compute different "concepts" like the territory assessments, etc., that lightvector is referencing.

_________________
be immersed

Maybe something "simpler" would consist in exploring a database of millions of selfplay games, and find local patterns which are frequently occurring. We would thus get a dictionary of life and death problems, shapes, tesujis, josekis, invasion techniques, sabaki techniques, etc. I expect that dictionary to contain thousands, or perhaps tens of thousands of patterns.

Then, if the student fails at finding the correct move in a given pattern, the AI could show sample positions containing analogue patterns and ask to "find the next move".

That's very hypothetical. I don't know how useful it would be in practice. Perhaps humans learn better with words than with patterns, so that human teaching cannot be replaced by machine teaching.

Reminds me of daily joseki: https://dailyjoseki.com/browse

_________________
be immersed

I was trying to address this in my post. The neural network must necessarily already have functions that approximately correspond to the strategies and techniques of Go. Exposing the underlying functions and activations may not be helpful because they are often intertwined. But there is research on identifying the top contributing features of the input data (board position) that lead to the result. Even coming up with an algorithm to explain the top contributing features would be difficult and the achievement might not even be as good as playing out variations with a game-playing AI. And all this is certainly not "teaching," which is what a lot of people would want.

I imagine a system could be designed to classify board positions using Go terminology. Something similar to Fairbairn's Go Wisdom format. The main issue is that you would need thousands (hundreds of thousands?) of these labeled positions. This is why Google has millions of users identifying buses and traffic signs (let's prove that we are human by training the bots).

So then, if one model predicts a move or sequence from an input (board position), another algorithm can determine the top contributing features (parts) of that input that caused that particular move/sequence to be recommended. Then different machine learning models can be used classify the input with the labeled human Go terms. And if the features contributing to the recommended move also contributed to certain classifications, then the reason that the move was selected could be based on those terms. The features could be represented along with their label. But even this is not "teaching." Perhaps more can be done.

----------

I think we are more like to have development efforts spent making secretly strong AIs to be studied by select top pros (is this not already a thing?) rather than have AIs developed for teaching the masses.

Well my take on the overall question is very clear,it doesnt make us lazy for sure not,because most western players are lazy from the beginning,they dont even try to find the best way to improve they rather try to fin something enjoyable which is totally fine, go here is something fun, so the whole argument of ryan falls apart since the initial and most important motivation of a go player is not improving but enjoying,the enjoying part can be helped by am

I had a hunch AI was doing more than just minimaxing - that's why it's AI and not some brute force or pure Monte Carlo - but lightvector's post is just stunning. And given the existence of that post, we are already closer to understanding/translating AI thinking. It takes the likes of lightvector to do so but an example now exists of a human clearly seeing AI developing a particular concept and being able to articulate it to a wider audience. True, the concept is a basic one, but it means there is discernible conceptualization. We don't know yet how much of conceptualization is going on - we have not discerned beyond the basic concepts.

When you say the "techniques of Go", do you mean techniques that humans have already established and/or understood? Or does this definition also allow for currently unknown-to-human processes which we could benefit from?

I suspect it's possible find the latter.

_________________
be immersed

When I said "techniques" I meant a technique (strategy) for judging a position to determine where to play. I did not mean technique as in tesuji or joseki. Surely AI might discover a new tesuji and they have their own joseki.

There is a reason why Deep Mind set Go as their initial goal -- it's super simple. There are only 19x19 inputs with 1 of 3 states at a fixed moment in time and a precise way to judge the result. I'm not suggesting that AI could not discover some new tesuji, because surely it could. But this still just be a variation on what is already known to humans. But it's not going to discover some unknown concept. Go is more difficult than other games but it's simpler than using machine learning to driving a truck or translate language.

What about an even more complex AI, like predicting movement of birds for focusing a camera from 1920x1080 image data with hundreds or thousands of different states for each pixel at 60 frames a second. Even in this more complex scenario there is no unknown to human technique that could be taught to humans so that they might better predict where a bird is going to fly. A computer might be quicker to recognize change in velocity than a human, but prediction relying on computation is not going to help a human. There is little use for a human to learn how to make assessments like an AI when they do not have the computation capability of AI.

Not sure if you're being serious but Ryan seems to be specifically talking about people who are not lazy, who are motivated to improved and are using AI to do so. Being unmotivated to improve is outside of Ryan's argument.

Ryan is not addressing the lazy player who enjoys learning about the ancient Chinese "Gold Scissor" joseki and "Jade Balustrade" opening while tracking what the Japanese tournament venues are serving for lunch in between deep foreign language search engines dives for clues about an illustrious wood collector.

...The venue for game 7 of the Meijin match didn't share their lunch menu during the stream this time. I did get to see video of a coastal train journey though.


Update: found it on Twitter. Iyama got soba again.

I don't understand why.

To give an example, if humans never understood what was meant by "thickness", but we saw that at some intermediate layer in a neural network that the AI seemed to be optimizing for what we call thickness, it might be possible to discover, "Hey - the AI seems to be optimizing for something here at this layer in the network. What is it? Is it there some useful pattern to find here?".

No?

_________________
be immersed

I think the bottom line is that a deep NN is needed to achieve decent results and deep NNs are not going to have any useful mechanism for a human to learn for game play. The human neural network is too different and too bad at computation. Plus, I don't think it's better than learning from trial and error with the game-play AIs.

But again, I think the reverse is helpful: having models to classify human game concepts, determining the top features that decided where the AI will play, and matching between the two.