Robots Invent Their Own Spoken Language | Hacker News Reader

Robots Invent Their Own Spoken Language

spectrum.ieee.org

127 pointseguizzo35 comments

or: robots use pre-programmed algorithm to assign randomly generated tags to emergent regions.

the cool thing here isn't "inventing a language"; it's that the system stabilises with different areas having different, consistent labels. the algorithm used to do that is based on social interactions. and this is cool.

"language" comes into it only in that the randomly-generated labels are encoded using phonemes and communicated via speaker/microphone.

it's sad that, when there is something cool, people still feel the need to introduce a gimmick.

[edit: it just occurred to me that maybe the cool bit has already been done by someone else, and all this work does is add the sound-based communication; i hope not!]

I mean, that is pretty much what language is. Learning of meaning through social interactions...

well, that depends on whether you worship chomsky or wittgenstein :o) although this is more like skinner :o(

dropping the pseudo-intelectual bullshit: there's no grammar here. yes, this is like saying "hammer" while holding a hammer aloft. but it stops there. language is more than a collection of nouns; there is structure. and any structure here is not learnt.

I think it's pretty cool that they constrained the robots to using human phonemes, because the investigators can actually listen and observe while the robots are talking. It might be more scientifically rigorous to prevent that kind of informal observation when testing a hypothesis, but they might learn something this way that will generate new ideas and new hypotheses to test.

This kind of research is more exploratory engineering than it is science. As someone who does this sort of thing [1], we don't really have a hypothesis and are looking for data to support that hypothesis. We have an idea - "Hey! I think we can get this cool thing to work!" - and then we grind away at implementing it until what we have is cool enough to show off. This is fundamentally different than pure science because we're not actually testing a hypothesis or a theory. (If you're thinking "What about the hypothesis, it's possible to do X?" well, that's a pretty weak hypothesis.)

Yeah, we have experiments, but those are used to show that technique X is better than technique Y when using metric Z. We're not improving our fundamental understanding of how how the universe works, we're improving our ability to build things - either by demonstrating a new technique to achieve an old thing, or by demonstrating that an entirely new thing is possible.

[1] Not robots or machine learning, but computer science research in general. I do high performance computing and systems research.

You know, before reading your post I never even questioned that I had a strongly scientific leaning in my mentality. Now I realize I'm much more of an exploratory engineer than a scientist. Actually, now that I think about it, haven't there been a great number of scientific discoveries that were, according to your distinction, actually exploratory engineering? I would think that following all the side-effects and byproducts of rigorous scientific study would qualify as exploratory engineering, wouldn't they?

Anyway, interesting perspective. Particularly on point, as I just finished reading Cory Doctorow's "Makers".

I've put a lot of thought into it, since it's what I do. I never felt comfortable calling myself a "scientist" because I was doing more inventing than discovering. I think there are cases where the distinction between the two can be fuzzy, but I think it comes down to have you learned something about the universe and the things in it, or have you learned a better way to build something? Most of the time I've learned a better way to build something. CS theory people can claim more fundamental knowledge, but they fall more in the math camp than science.

And make no mistake, I would be flattered to be called a scientist. But I have friends in physics and biochemistry. I know that there is a fundamental difference in our research.

However, knowing how to design an experiment that will show what you want to show is required for both science and engineering research.

It's still interesting. More interesting: set up robots that can gradually associate particular movements with particular sounds. That could lead to a truly organic (and self-developed) robot language.

Thanks. I knew reading the comments before excitedly RTFA’ing was a good idea.

Grrrr. No, they haven't invented a language - they have invented a vocabulary. The grammar seems to have been predefined for them.

There is a school of thought[1] that believes, that to a certain extent, this is true of natural human languages too! Some rules of grammar might just be hard-wired into the brain and not learnt as we grow up.

[1] http://en.wikipedia.org/wiki/Universal_grammar

From that article: "Since their inception, universal grammar theories have been subjected to vocal and sustained criticism. In recent years, with the advent of more sophisticated brands of computational modeling and more innovative approaches to the study of language acquisition, these criticisms have multiplied."

It seems likely that there are properties of the human brain and body which would lead to regularities in human languages. However, I believe that the proponents of universal grammar are overstating the case.

whenever I hear about robots or software agents "inventing" or "evolving" some aspect of human intelligence, I'm always skeptical because it sounds like they were programmed such that this outcome is almost pre-ordained. for example, evolutionary "art" that "draws" the mona lisa - I saw an example of this, but the actual mona lisa was essentially the objective function of an optimization, so why wouldn't it draw it after enough steps? granted this article doesn't say much about implementation, but it sounds like they were programmed with the interpretation of words being associated with locations as a given.

I admit I'm a novice in this particular domain so I can't speak too intelligently about it, but am I missing something?

'Approximate that using this limited tool set and tell us when you think you're close' is a nontrivial problem. Without getting into a discussion of what creativity 'is', we can probably agree most everything about human activity begins as imitation.

And humans, already having an innate intelligence, can easily imitate something. Without human interference a robot would rust to pieces before it ever moved a millimeter (we do have to tell it what to do).

We could invent synthetic creativity, but it in my opinion it'd never progress past its synthetic nature.

Why can’t humans create what nature created? Creativity is not magic, I see no reason why it should be in principle impossible for humans to build creative systems.

Let's start with building another human, then, beginning with elemental chemicals. I think we all know how far we can get with that.

Or, let's try to engineer an entire economy to supplant the emergent system of markets. Oh, wait, that results in millions of people starving to death.

Some systems are so complex that it's beyond our abilities to dream about how we might approach such a problem, let alone execute it. Given that we've been so unsuccessful in assigning a definition to creativity or intelligence, I think that these things might be in the same category of overly-complex systems.

If evolution can do it, we will be able to do it, too. (Given enough time.) I’m not at all sure how markets figure into this. Human brains are not markets, this comparison strikes me as non-sensical.

I don’t see you claiming that there is anything in principle impossible about that.

brains and markets are both complex adaptive systems. they aren't similar in high-level function, but they are with respect to low level dynamics.

You forgot: Heavier-than-air flying machines are impossible. Oh wait...

Edit: Not sure why I got downvoted, probably the snarkyness. Fine, I can also put it the long-winded way:

There used to be a time when flying seemed beyond our abilities to dream about how we might approach it. Likewise for most other achievements that we have made.

What's going to stop us from building a human body, including the brain, once we have the manufacturing technology to do so?

>What's going to stop us from building a human body, including the brain, once we have the manufacturing technology to do so?

it's that we don't know what "manufacturing technology" we need, because we don't have a deep enough understanding of how the brain works. I think the parent's point was that attempts to engineer complex systems fail because these attempts can't reproduce the emergent behavior that arises from many layers of interactions from individual agents. a plane is complex, but it's not a complex system.

edit: I actually disagree a bit with myself - you can make a case that planes have elements of complex systems, but it's not at the same scale as a brain, in my opinion.

Yes, I don't dispute that we can not build a brain today. But I'd say this is not very different to our inability of building a plane (or understanding how it could possibly work) back in the 1900s.

I just checked wikipedia and about 100 years ago (February 1911):

    The first official air mail flight takes place from Allahabad, India to Naini,     
    India, when Henri Pequet carries 6,500 letters a distance of 13 km.

Today, 100 years later, I tell you about this event by typing my "letter" into a portable device that will soon transmit my message over the shy distance of ~8000 miles, in under 100 milliseconds.

So, looking back at what we've achieved in the past 100 years, and unless someone discovers some inherent property of the human brain that prevents it from being replicated (did God add DRM?), I remain pretty optimistic that we will be able to build one soon.

No one person can invent robotic creativity, therefore a team is necessary. If we presume that the team is comprised of members renowned the world over for their robotics and programming work, and are leaders in their field it doesn't guarantee success (even with technology 200 years from now).

We cannot even agree what creativity is and where/how it comes from the human brain. Without a full understanding of creativity, which I'm sure a few bus-loads of philosophers and psychologists could never agree upon, that is communicated without flaws to an expert robotics team... I hope you see the ridiculousness of this idea.

For all our understanding and consensus on the matter, creativity might as well be magic. And, as much as I've programmed and studied theory... I still don't know how to implement magic.

Unless, of course, Ruby or Python already have "creativity" metaprogramming?

    class Robot
      include Creativity

      creative :language, :syllables => 2, :syllabary => %w{ ja ya ri je pi ze ku zo re }
    end

It's a hard problem, that's all. Not in principle impossible.

Then we'll have to agree to disagree.

I agree that the principle is possible because I view almost every principle as possible. But I think the achievement of said principle is far beyond man's capabilities. And I think the primary limiter is us, that we cannot produce something greater than our own intelligence. (I have seen many try, and fail, to build something that reaches even their own intelligence.)

This isn't a language: these are just randomized names. Anything can do that, it just takes the place of numbers, hashes, or whatever database you'd otherwise use.

Come back when they (re)invent nouns, verbs, adjectives, etc. Like "beware that pipe", only "for 'droids".

It's like programmers using plaintext for a protocol instead of binary and getting the front page of Hacker News.

Cool, except that humans don't always invent words from random syllables. This is symbology, not language. Words often represent the sound of an action or are derived from mutating or combining other words for related concepts or names.

Automated Not Invented Here Syndrome (ANIHS)

Have these robots invented words for what's outside of the space they can sense? That would be interesting.

Wouldn't it be much better if we seek a robotic natural language or modify existing natural language to make robotic ?

Looks like finally we can have a real C3PO

It's good to know that the future Skynet will be able to talk to each other and we won't have a clue what they are on about!

Wow, can no one take a joke on here

Only if it contributes.

Reply on news.ycombinator.com