Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

I'm Scott Aaronson, quantum computing/computational complexity researcher. AMA

news.ycombinator.com

Post reply on HN
Print view View original

321–330 of 373 posts

Re: I'm Scott Aaronson, quantum computing/computational complexity researcher. AMA

#321

Post by ScottAaronson »

ScottAaronson wrote: post #307
YeGoblynQueenne wrote: post #262

Earlier quoted context omitted.

>> where indeed they use machine learning to induce the rules of chess from a large number of played games (and then learn to play better than any human). They didn't induce the rules of chess! What they did was learn an evaluation function for chess moves, from scratch, i.e. without giving it the rules or any hand-crafted features. So they learned a classifier for good/bad board states, which was later used in an al…

Aha, thanks for that extremely useful clarification! I was also confused from the abstract and intro about exactly what the inputs and outputs were. Yes, the clear impression was created that they learned the rules---in the sense that they could then generate moves that would (usually? almost always?) obey the rules---when it sounds like really they just learned an evaluation function that encodes lots of implicit kn…

On further thought, we should distinguish two problems:

(1) "Induce" the rules, in the sense that after training on millions of played games, you now have a neural net or whatever that plays mostly or entirely according to the rules even though it can't articulate them.

(2) Output an explicit description of the rules.

I could easily believe that (2) is beyond the current abilities of AI, even if it turns out that (1) is doable.

Re: I'm Scott Aaronson, quantum computing/computational complexity researcher. AMA

#322

Post by your-nanny »

ejanus wrote: post #309
stcredzero wrote: post #190

Earlier quoted context omitted.

Here's my attempt at a two sentence over-over-over simplification that at least gets people away from the "magic bit-sting that contains your answer." (It also harkens back to an old Einstein quote, so may be attractive to science writers.) Quantum computing is a technique that lets you sample a problem's answer-space using "loaded dice," such that the problem's correct answers correspond with probability spikes in y…

"Loaded dice ", I have not heard that before... Is it same as Biased dice

Yes

Re: I'm Scott Aaronson, quantum computing/computational complexity researcher. AMA

#323

Post by lisper »

Y_Y wrote: post #299
lisper wrote: post #138

Earlier quoted context omitted.

> Who said anything about physics? You did. You asked about why the universe we live in has a particular property. That's a question about physics.

Well I obviously don't think so, and all you've done is state the opposite claim.

What do you think the word "physics" means?

Re: I'm Scott Aaronson, quantum computing/computational complexity researcher. AMA

#324

Post by YeGoblynQueenne »

ScottAaronson wrote: post #321
ScottAaronson wrote: post #307

Earlier quoted context omitted.

Aha, thanks for that extremely useful clarification! I was also confused from the abstract and intro about exactly what the inputs and outputs were. Yes, the clear impression was created that they learned the rules---in the sense that they could then generate moves that would (usually? almost always?) obey the rules---when it sounds like really they just learned an evaluation function that encodes lots of implicit kn…

On further thought, we should distinguish two problems: (1) "Induce" the rules, in the sense that after training on millions of played games, you now have a neural net or whatever that plays mostly or entirely according to the rules even though it can't articulate them. (2) Output an explicit description of the rules. I could easily believe that (2) is beyond the current abilities of AI, even if it turns out that (1)…

>> Since you obviously know this subject, what are the best current results in the direction of inducing the rules of chess from played games?

To be honest, I don't think there's much work on inducing the rules of chess, in particular. It's probably considered a) easy enough to do by hand and b) too hard to machine-learn.

>> On further thought, we should distinguish two problems:

(1) - Yep. The most likely approach would be a classifier trained to label moves as legal/ illegal. The resulting model would be a vector of numerical parameters so not a traditional rule base. It would also only be correct within some margin of error, probably not 0, limiting its uses (e.g. it wouldn't make sense to train it to play and then pit it against a player with a correct rulebase; they wouldn't be playing the same game).

>> I could easily believe that (2) is beyond the current abilities of AI, even if it turns out that (1) is doable.

Also yes. Exactly on point in fact.

When we're talking about learning rules, we 're talking about learning automata, the subject of inductive inference, an older branch of machine learning (well, ish) that fell out of favour after a bunch of theoretical results showed it was basically impossible to learn any interesting class of automata from examples (the most famous is Mark E. Gold's result from Language Identification in the Limit; only finite languages can be learned from finite examples, anything else is only learnable "in the limit", from infinitely many examples, or an all-knowning oracle, etc).

Modern machine learning starts with Valiant's A Theory of the Learnable, which introduced PAC learning and a relaxation in the assumptions of inductive inference, about what should (and, therefore, can) be learned.

In short, the difference is that inductive inference attempted to learn complete definitions of various automata, whereas modern machine learning attempts to approximate them; well, strictly speaking it's about approximating functions, not automata as such.

So yeah, pretty much, like you say: (2) is, in principle, not possible whereas (1) might even be possible in practice.

Now, normally this is where I'd plug my own research, but I've already written plenty :)

Re: I'm Scott Aaronson, quantum computing/computational complexity researcher. AMA

#325

Post by ScottAaronson »

tjhance7 wrote: post #162
ScottAaronson wrote: post #88

Earlier quoted context omitted.

My position on the technical fundamentals never changed much: namely, D-Wave is building devices that could be interesting from various engineering perspectives, but that as far as most of us can tell, are not getting speedups over existing computers that are clearly attributable to quantum computation (as opposed to building special-purpose hardware that's, essentially, very fast at simulating itself). If you want q…

> This is partly because, even assuming you accept the Many-Worlds Interpretation, "your" moral concern and responsibility presumably extend only to those branches that are in "your" future -- you have no contact with the other branches! Would you say that the only moral way to implement quantum suicide is with a Doomsday Device that would destroy the entire world, thus ensuring your actions won't affect anybody else…

No, I wouldn't recommend that either. :-)

Re: I'm Scott Aaronson, quantum computing/computational complexity researcher. AMA

#326

Post by ScottAaronson »

BenoitEssiambre wrote: post #229
ScottAaronson wrote: post #15

Earlier quoted context omitted.

Sorry, I don't know what "emergence" means in this context. If you mean AGI, I think (hope?) that we're still quite some ways away from it. Yes, quantum computing could help with AI -- for example through Grover's algorithm, which lets you solve many search, optimization, and planning problems in roughly the square root of the number of steps you would need classically. But it's a complicated story: many of the probl…

"not to mention our own brains" Is this a known known that our brain doesn't use QC?

No, it's an unknown known. :-)

Re: I'm Scott Aaronson, quantum computing/computational complexity researcher. AMA

#327

Post by ScottAaronson »

n0mi1k wrote: post #17

If advanced quantum computing were available in smartphones and other small devices today, what applications would be possible that are not currently? (Or which would be greatly improved) Would it affect the things many people do most on their phones like messaging, news and social media

As far as I know, it would have zero effect on any of that ... unless you need to simulate quantum physics or chemistry, factor large integers, calculate discrete logarithms, or possibly solve some large optimization problems while doing your messaging, news, and social media. :-)

Re: I'm Scott Aaronson, quantum computing/computational complexity researcher. AMA

#328

Post by ScottAaronson »

noobermin wrote: post #225

Hey Scott, thanks for doing this. What do you think of the Copenhagen Interpretation especially in light of the delayed choice quantum eraser experiment[0]? I know interpretation of QM is something actual working scientists try to avoid (I know I do such for my little corner of physics) but given how much interpretation informs intuition it's worth considering once in a while. [0] https://en.wikipedia.org/wiki/Delaye…

I confess I get annoyed when people make arguments about the interpretation of quantum mechanics using complicated thought experiments involving lasers and mirrors. Why not just talk about it in the modern way, in terms of qubits and states? :-)

For some of my recent thoughts about interpretation of QM, see this blog post, and especially the discussion in the comments section: https://www.scottaaronson.com/blog/?p=3628

Re: I'm Scott Aaronson, quantum computing/computational complexity researcher. AMA

#330

Post by YeGoblynQueenne »

YeGoblynQueenne wrote: post #324
ScottAaronson wrote: post #321

Earlier quoted context omitted.

On further thought, we should distinguish two problems: (1) "Induce" the rules, in the sense that after training on millions of played games, you now have a neural net or whatever that plays mostly or entirely according to the rules even though it can't articulate them. (2) Output an explicit description of the rules. I could easily believe that (2) is beyond the current abilities of AI, even if it turns out that (1)…

>> Since you obviously know this subject, what are the best current results in the direction of inducing the rules of chess from played games? To be honest, I don't think there's much work on inducing the rules of chess, in particular. It's probably considered a) easy enough to do by hand and b) too hard to machine-learn. >> On further thought, we should distinguish two problems: (1) - Yep. The most likely approach w…

Dah. Hang on, what am I talking about- chess has a finite number of moves so the "language" is finite. That means it's learnable in polynomial time, from finite examples ... in principle :)
Post reply on HN