Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[Reelya]

Those people aren't posting in this thread, and the relevant posts were in an unrelated thread about chatbots. It's a bit patronizing to claim you're talking about some unrelated people after you acted like you were slamming the people posting here.

Also I think the "magic" is on the other side. The main claim I made is that some things that others claim are unknowably human are in fact resolvable through machine learning and statistics methods.

It's the people claiming "fun" is this unknowable special snowflake stuff that only humans can design for who believe in magic.

[Reelya]

I've just been responding to your points all along. You're the one who tried to hack apart what I said, constantly. Now that I started pointing out what you've been doing you're getting upset about it.

I mean, you hacked apart the video itisnotlogical linked, and said "that just proves my point that XYZ is all bullshit".

Also, that video just goes to show my point- the spread of performance across each game is ridiculously wide, from well-performing on Gradius to being barely able to even play Megaman. It barely even knows how to move in Megaman.

But which point that you made does this relate to? You were hacking apart my idea that machine learning could be used to tune how engaging an arcade game jump was. And now your claiming to have been defending an entirely unrelated point about how you can't create general-purpose AIs? Who was arguing for that? There, you created a whole new argument about something nobody was arguing about.

Previous to the shown video:

The main issue with machine learning is that algorithms trained for one task tend to suck at any other seemingly-related tasks. You could train a neural net to beat a single level of a Mario game, but I doubt you could make the same neural net beat every single level in the whole game.

The guy 100% proved you wrong, about a statement you made which was 100% assertion and not related to any previous point. You then shift the goalposts when you were proven wrong, and do ad hominems such as accusing people of supporting "magic". Now you want to play the victim.

[Reelya]

Well the generalization of a neural network is a brain. So actually I'd argue that these are scale and connectivity limitations, and not underlying limitations of the neural approach.

When people talk about the limits of what an NN can do, they are talking about feed-foward layer-based networks (which don't have any loops or time variables involved). Those are not the same as saying neural networks in general, which could have feedback loops and larger structures, as well as memory circuits.

Additionally, genetic algorithms can be turing complete, those are limited by the limits of what is computable. Then that comes down to whether you think human IQ is something beyond turing complete problems and that humans have "magic stuff" in their heads which is beyond computation.

If you want to talk limits.

[Reelya]

Sorry but at this point I'll have to ask you to make a new thread if you want to talk philosophy.

This is specifically a thread for applied science and technical solutions to existing problems. If you have some ideas on a thing you actually want to build you can discuss them here. It's veered into some machine learning, but that's only OK as long as we're discussing the application of real existing systems and ideas, to specific problems. Slamming down people's ideas based on philosophical concepts or a general "it can't be done!" attitude is against this thread's reason for existence.

[itisnotlogical]

As for genetic algorithms, they're really useful for incremental improvements towards a solution, but they have a real issue with getting trapped in ruts or local optima.

That's what calculus is for. Of course it's more complicated than that answer indicates, but it's not like there's no way to mathematically pick an optimal solution given a function that represents the relationship between some number of properties.

[wierd]

which you can only prove after exhaustive elimination of the complete solution space... At which point, your genetic algo is not really any better than brute force, aside from reaching a useful local optima sooner.

[Reelya]

This is part of the reason that they prefer mutating the bits in the values rather than e.g. adjusting the float values. If you mutate bits directly, then you get big jumps as well as small jumps, so you can scout out a larger search space, instead of just iteratively optimizing to a local minimum. Big jumps massively increase the chance of getting into the "zone" of a different minimum.

Also if you could infer any meta-characteristics of your search space then you could do a much more thorough search. e.g. say your wanted to optimize "find the highest point on Earth" you'd be silly to go over the space millimeter by millimeter, just in case you missed some really thin mountain somewhere. Or if you just start on a random point and walk uphill and expect that to work, that would also be dumb. Nope, you'd sample evenly spread out points on the earth, then you would pick likely points and do a more fine-grained search around those.

[wierd]

Better understanding of the solution space, and the mechanics of the search, can systematically exclude huge sections of the solution space as well, for being absurd to even attempt seeking solution in.

This is one of the ways some of the more interesting attacks against modern cryptography work.  With the right understanding of the problem, you can prove that certain sets of solutions are not valid-- EG, "we know it is an odd number" which would exclude all even valued solutions to a number search space-- etc.

[McTraveller]

I can't figure out how, if http://www.earthharpsymphony.com is actually indeed an acoustic instrument as claimed, how you can get audible frequencies out of strings that long without insane tension in them - which they don't appear to have based on the images (especially with the one set up on a mountain!?).

I suspect that it's just that the strings set up excitations in the resonator on the stage, and the resonators are the actual source of the audible sounds.

A pretty cool engineering / art mix though!

[wierd]

Looks like instead of strings/wires, they are actually metal rods. In some of the videos I see, (like the america's got talent video), there appear to be toppers on the end.

My guess is that it is actually a stiff metal rod that has a resonant characteristic to leverage self-polyphony of the rod. This would be improved by having a hollow rod, as then the cavity inside the rod would serve as a conduction channel. Depending on where he is holding the rod, the "short side" near the base of the instrument would be the sounding portion, with the long side being only a sympathetic resonance source. If the lengths of the rods are properly computed, self-reinforcing waveforms would propagate in the long end of the rod, from the friction induced resonance of the short end of the rod, through stimulation carried over the air channel inside, via sympathetic resonance.

Just a guess though. It would not need to be so tight, if it was instead stiff.

[Starver]

The frequency is proportional to the harmonic you're looking for (even if the main harmonic is too 'infra' for you, which it's possible) and the root is proportional to the tension (so too tense and it would be too high for usefulness) and inversely proportional to the length (very long, so very low before you go for harmonics) and root of the string density.

The tension is restrained by the same engineering practicalities that brought you to a 'string' of a given material and density, and (safe) tension only serves to somewhat shift the fundamental harmonic, whilst the 'flavour' of the note could easily be from whichever harmonics we hear rather than could only 'feel'.

At a guess, that's the effect being used.


(Ninjaed, and I couldn't get  much detail from the site, seems to want to display pictures on the RHS, but all I'm getting there is "click to see more" on occqsional 'frames' of background colour that happily scroll but show nothing...)

[Reelya]

...what are you even talking about?

You were the one who brought up the brain part of neural networks before you edited your post. What are you even going on about?


Do you just have a problem with me personally? I'm so confused.

I was just pointing out that getting into discussions about philosophy is too far from what this thread is about. it's for posting tech news. Specifically, news about current applications of technology to solving problems. We were starting to veer into philosophy, and I think that's way too far off course and should have it's own thread. There have been threads about philosophy of mind and hypothetical artificial intelligence before, but doing that again here only dilutes the current topic.

We could have 20 pages of philosowank about the limits of AI without anyone agreeing on anything. I don't think that is beneficial, and it would effectively swamp any actual discussion of tech news under the weight of hypotheticals.

I don't hate you, but you did use an argument technique which consisted of a series of non-sequiter complaints to every little thing wrote (and those individual complaints were almost all either tangential or just plain wrong), when I was pretty clear about what I was proposing the whole time, then you seem to have done some strawmanning and bringing up targeted complaints and when i challenged whether those complaints were valid, you said "i meant people in general not people in this thread". If you meant people not in this thread, don't bring it up in a context that sounds like you're criticizing people here.

[Reelya]

Nobody was making the counter claim however. Remember all I said was that genetic algorithms could be used to tune a single arcade game challenge.

Just on the last point you've been constantly shifting your goalposts:

The main issue with machine learning is that algorithms trained for one task tend to suck at any other seemingly-related tasks. You could train a neural net to beat a single level of a Mario game, but I doubt you could make the same neural net beat every single level in the whole game.

But then an AI was shown which can play dozens of different games. It can do great at some, not so great at others. But it's the same AI in each case. The mere existence of such an AI disproved your clearly-stated point that an AI couldn't be made that can beat more than a single level of Mario.

[Reelya]

I was responding to your line of conversation when I wrote that.

Now I'm saying we've gone too far off the thread topic.

But I will point out that you shifted the goalposts on your "an AI couldn't beat all levels of mario" argument. Because when the existence of an AI was shown that beats multiple different NES games, you suddenly declared that it wasn't sufficient because it can't beat all NES games.

This was clearly shifting the goalposts from your previous statement that an AI couldn't beat more than a single level of any one game.

[Reelya]

I never made threats to throw you out of the thread. I said to bring it back on topic. And that if you want to talk philosophy then it needs it's own thread. That is not a threat to throw you out of the thread.

And the accusations are just pointing out the exact things you did.