AI names colors much as humans do | Hacker News Reader

AI names colors much as humans do

ai.facebook.com

22 pointsAnon8415 comments

I definitely don’t understand this paper to be what I thought the headline meant.

I thought the headline meant that generally AI systems would end up having color systems which distinguished blue and orange more readily than taupe and beige.

But the paper doesn’t seem to talk about that, except in the abstract - the paper instead seems to report that the more complex the ‘allowed’ grammar, the more specificity AI can have about colors. And, by the way, that’s true for human languages too, and you can train a grammar that’s close to the info-theoretic limit, which by the way is what you can say about most human color languages.

It seems to me like you could train an AI to report hexadecimal for colors pretty easily, so I remain confused about a lot of details, most of which seem to be embedded in the referenced paper “Efficient compression in color naming and its evolution.” I haven’t read that paper. Without reading it, I’m not sure if the networks developed just carried over concepts from it, and thereby got similar segmentations (and the graphs look kind of similar to me, but not exact or anything), or if there is something fundamental.

All that said, it would be REALLY cool if alien intelligence-like assessment of colors saw things the way humans see them, that would be rad. I just have no idea if this research proves that or not.

Not sure what to take of it. At first, I assumed that they mean that NN learned to divide color plane similarly to how humans do, without learning on human provided data, which would be absolutely incredible, but no, no wonders here, the actual vocabulary was pre-defined. The actual result is that when you task NN with making up discrete words (and being discrete is pretty much the defining feature of any human language, quite naturally) and measure their "efficiency" by some information-theoretic metric, you get a distribution similar to what human languages have. And, again, that could have been interesting, except it turns out to be right on the border of what is "theoretical limit". Well then, if there is only 1 theoretical limit, so it's no big surprise that when you train the system like that, penalizing inefficiency by any related metric (let alone the same one), you'll get the same distribution.

I'm cautious not to say that it has nothing to do with linguistics, because half of linguistics is math because of the nature of what is language. But that's about it. That is like setting up an actual physical experiment with the Galton Board, to "discover" that you'll get a normal-distribution looking pattern with aluminium balls the same way as you'll get with wooden balls. That's all about math involved in some information-theoretical concept, not about human laguages per se.

Trained on human data

This doesn't seem to be correct. The article talks about reinforcement learning agents optimizing a communication game to trade off color description accurately vs. effort.

The words they use aren't real words, they're partitions of the color space, and the researchers found that the partitions the agents came up with to win the game were similar to human partitioning of the color space.

Now, did the design of the game and the reward function smuggle in human notions of reasonableness that made the outcome a foregone conclusion? Maybe that's a more reasonable criticism, I don't know.

The study relies on "colors" defined by human perception, which could be interpreted as a form of training, when all inputs are restricted to that definition.

The efficiency/complexity insight doesn't rely on that data, but the human-like output produced by human-like color data combined with communications limitations does rely on it, and that's what the article is all about.

AI intended to replicate human behavior replicates human behavior

Edit: It’s like domesticating wolves for the purpose of training them to run around an obstacle course at Westminster to showcase what it would be like if they were still wolves

A quick read of the article leads me to believe that the AIs were inventing language tokens of their own. What makes you think it was trained on human data?

Seems so, but that's less important than the title suggests. The paper is far beyond me (I can't even figure out what the "IB plane" is) but the key insight is how complexity tracks with communications efficiency when their models make their own communication methods between each other.

The article summarizing it underplays that part.

Could you elaborate?

Suppose they trained it on data from dogs, how would the result be different?

I believe the parent was asking in what sense these models were trained on human data.

If I decide to throw in the towel in my business and go and become a well-paid AI researcher at some big company, which companies currently have anything like Facebook's AI team or Google Mind, but aren't a) in the Bay Area, b) FB or Google? Netflix seem to have a decent AI team, although they're in the Bay Area.

IBM

So how many times did they have to tweak their algorithm to get this nice result?

Only the communications limitations changed, according to the paper.

>As we relax discreteness, the emergent naming systems become complex beyond what is attested in human language, and, eventually, significantly inefficient.

Reply on news.ycombinator.com