Hacker News Reader

Comment by hasenj | Hacker News Reader

Actually this AI kind of confirms these myths, since its basis are not in mathematics but in neural networks. While it could be argued to be just math, so is the brain, but that's besides the point. The point is, even the programmers have no idea what the AI is thinking.

The way it picks moves is very similar to how top professionals do.

Intuition is reduced to memories stored vaguely as neural connections.

View on HN

I fear this is an overly mystical misinterpretation of neural nets. (Standard feed-forward) neural nets are layers of non-linear feature transformations. They take inputs and at each step transform those inputs into a more compact representation, distilling the inputs into their most important factors (and throwing away unimportant factors).

So the most likely explanation is that policy/value nets in AlphaGo have learned to extract - with cold logic - the key factors that make up what humans believe to be "good" board positions.

It has little to do with voodoo about neural connections and magic emerging from the weights. AlphaGo has most likely managed to identify the important factors of good board positions (by seeing tons of examples of good and bad moves/positions). It only appears to be magical because these factors are most likely very complex and inter-dependent.

This is supported by the AlphaGo paper - they report that AlphaGo without tree search is about as good as the best tree search programs (amateur pro level). So AlphaGo has taken amateur-pro-level board analysis ability, and combined it with tree search, to achieve top-player performance..

> So AlphaGo has taken amateur-pro-level board analysis ability, and combined it with tree search

I don't think that's an entirely accurate way of putting it, because a player who reaches that level is also doing a little tree searching. Maybe if you found some human who managed to reach "amateur pro" level by playing purely on snap, instinctual decisions without any logic or exploration of variants at all, yes, then you could say their ability to evaluate positions is as good as AlphaGo's.

But I would guess that you are right anyway that we can deduce that its ability to evaluate a board position really is below that of better professionals, and its huge strength is due to the tree search (which of course involves a second "policy" net to pick moves to explore).

In go, the tree searching is named "reading". Even beginners need to read very deep trees (more than 7 moves) to anticipate the result of a Semeai (capturing race).

> distilling the inputs into their most important factors (and throwing away unimportant factors).

Isn't that pretty much the definition of intuition? Combining a bunch of things in some unknown and nonlinear way to result in a 'feeling' about the situation?

Pretty much. The nature of evaluating a Go position, is that it has some notions of similarity like islands, "aliveness", and local features that can be looked at liberty-wise. But in order to fully understand how it all pieces together, one must have seen other positions to intuit the kind of future playstyle that will result in the game. Anyhow, it's perfect for a stochaistic system that just recognizes patterns. As long as you can provide the tree search (ie. the partial position evaluation) you can basically just let it rip.

Yes. But the person I was responding to was definitely up in the clouds describing things as "magic" (and implicitly equating intuition to magic).

Is that really different to what humans do, though?

I fear this is an overly mystical misinterpretation of neural nets.

Isn't most human experience an overly mystical interpretation of physics/chemistry?

Well obviously it's reduced to actual algorithms.

The point is, the programmers don't quite understand what exactly are those features that the neural net is seeing.

You can still beat Go using "simpler" math. If you have enough compute power, you can always just minimax the whole game tree. Neural networks aren't un-mathematical; they're just a slightly more complicated technique for discovering an approximation to a function that does what you want (even if you're not sure what that function looks like internally).

So AlphaGo's reign will last only until QuantumGo arrives on the scene? It would be sort of ironic to have spent decades developing practical AIs with classical computing to have them swept aside only when they started to really deliver results...

Stored discretely and permanently you mean..

I mean it's not clear in anyway "where" is the intuition, it just sort of magically emerges from the connections and the weights in a way that no one can really grasp. If nothing else, just due to the sheer number of neurones.

Is this just something you're imagining, or did Google's developers explicitly say, "AlphaGo is beyond our understanding. We have absolutely no idea how it makes its decisions"?

I'm pretty sure I heard them say this.

They can see certain stats and high level overview.

But they have no idea what it's thinking.

They know the general pattern of the algorithm. They even explain it.

But the algorithm involves two deep neural networks, and they don't really know what's going on inside them.

One of the developers showed up during the commentary on the second game and talk about this stuff:

https://youtu.be/l-GsfyVCBu0?t=2519

AlphaGo developers don't understand how it works in the same way you wouldn't understand how the program you've written to find prime numbers actually found a big prime number. The sequence of operations is known, but numbers are too big to be comprehended.

I think its more like, real parents dont understand why their children do the bizarre things that they do.

I don't think so.

Which kind of makes me think. Would machine learning succeed even mildly at recognizing primes? Would we be able to decode the final weights after weeks of learning, and find a sieve program encoded as data?

Well apparently my question leads to some deep mathematic theories about languages encoded by data. http://cstheory.stackexchange.com/questions/15039/why-can-ma...

Reply on news.ycombinator.com