Wikipedia:Reference desk/Archives/Computing/2019 July 25

Welcome to the Wikipedia Computing Reference Desk Archives
The page you are currently viewing is a transcluded archive page. While you can leave answers for any questions shown below, please ask new questions on one of the current reference desk pages.

July 25

Published literature about the limits of machine learning

What problems are non-tractable for a machine learning approach? Where should not even try a ML approach?

And specially, what problems are solvable, but not by an ML approach? C est moi anton (talk) 13:01, 25 July 2019 (UTC)

Well, computers can't think "outside the box", where "the box" is the range of options they are programmed to consider. For example, a computer programmed to find new ways to get clean water to a village where the well has run dry wouldn't be likely to consider the option of moving the village to where there is clean water. SinisterLefty (talk) 14:03, 25 July 2019 (UTC)

Historically, AI systems have been excellent at "thinking outside the box". Fractal antennae or evolved antennae for handheld microwave devices, or even mobile phones, are a good example, some of the best examples of which have been produced by evolutionary algorithms.

The reason is that "the box" is generally an emergent set of artifical constraints, based on widly-held but unfounded assumptions. The AI system just doesn't code the box, so isn't limited by it. Andy Dingley (talk) 16:43, 25 July 2019 (UTC)

That would be "inside the box", where it comes up with a design within the programmed constraints (physical dimensions, etc.). Now let's see it come up with a solution outside those constraints ("even though you said to limit the height to X, we could move component A to position B, thus making room for us to extend the antenna beyond X, and increase gain significantly"). Aside from sci-fi (like in WarGames where the computer learned that "the only winning move is not to play"), that doesn't happen.

Your example is a good use of the best of both the human mind and AI. That is, a human had to program it to ignore the traditional assumptions, on the suspicion that they may not be correct. The AI was then able to run billions of simulations to confirm that suspicion.SinisterLefty (talk) 16:59, 25 July 2019 (UTC)

I appreciate speculation about the topic, but I really mean I want published literature about application field/limits of machine learning (be it deep learning, genetic algorithms, or what you want to call it, when a computer is left all alone to discover the rules of something). C est moi anton (talk) 20:24, 25 July 2019 (UTC)

Here's an article on how bad input limits results: [1]. SinisterLefty (talk) 20:29, 25 July 2019 (UTC)

You might start with "The Book of Why" by Judea Pearl. He argues that the missing ingredient in the current neural net approach is causal inference. 73.93.153.154 (talk) 11:29, 26 July 2019 (UTC)

Hard theoretical limits are a subject of computational learning theory. One of the core results is that you cannot reliably learn an arbitrary computable function (nor even total computable function) from an initial segment of values (you can, on the other hand, learn a set thats isomorphic to the above - but with a trick that's not useful in practice). There are stronger paradigms, such as probably approximately correct learning that are more directly applicable to practical machine learning. One thing I found out in practice is that you cannot learn from data you don't have - when we build a diagnostic system for the diagnosis of poisoning, we got a long list of physiological measurements, but when we asked the physician how they usually detect alcohol poisoning, he told us "by sniffing at their breath" ;-). --Stephan Schulz (talk) 12:45, 26 July 2019 (UTC)

Perhaps breathalyzer tests should be incorporated into the diagnostic test battery. This would also compensate for doctors with a poor sense of smell, such as smokers. SinisterLefty (talk) 14:05, 26 July 2019 (UTC)

Yes, but much as I hate to cite a torture-supporting scumbag, there may be unknown unknowns. If your inputs omit something critical, your model will not perform well, and you may need more knowledge to figure out why. --Stephan Schulz (talk) 23:10, 26 July 2019 (UTC)

Agreed. SinisterLefty (talk) 00:08, 27 July 2019 (UTC)

I'm particularly worried that autonomous driving AI programs may lack important "out of the box" thinking. For example, a beach ball rolling across the street means "slow down because a child may dart into the street to retrieve the ball". Could it figure that out ? Perhaps, after running down several kids, it could. SinisterLefty (talk) 00:08, 27 July 2019 (UTC)

I do not believe autonomous cars react poorly compared to humans to some object dashing through the street in front of them. A human would maybe pay more attention when this happens, but a selfdriving car is already at 100% attention all the time. A human could slow down, or not. The car could slow down always when something unexpected is happening before its eyes (or before its sensors).

But I would be worried too, if only ML is being used to develop the system. The ball rolling onto the street is a pretty uncommon scenario, but with devastating consequences. Can we reliably expect that our data set will have this case covered? Will the algorithm be flexible enough? That is, would a skate instead of a ball, trigger the same braking pattern?

C est moi anton (talk) 12:19, 27 July 2019 (UTC)

It's not just about paying attention, speed also matters. At full speed, there may just not be time to avoid hitting a kid who darts out from between parked cars. Physics limits what can be done. So, you need to slow way down until clear of the area, even though the visible threat (the ball) is gone. SinisterLefty (talk) 20:43, 27 July 2019 (UTC)

Driving is littered with black swan events, which machine learning is not very good at. I presume that's probably the hardest problem for driving AI right now. 93.136.43.218 (talk) 22:03, 27 July 2019 (UTC)

Just to get all our ducks in a row, I assume you meant to link to black swan theory and not lead us afowl, on a wild goose chase, with your avian link. SinisterLefty (talk) 02:08, 31 July 2019 (UTC)

Interesting example. Won't comment on the morality of it but apparently you are NOT supposed to slam on the brakes if (say) a squirrel runs out onto the road, because the traffic hazard from slamming the brakes is supposedly worse than the consequence of running over a squirrel. I've heard of people failing driving tests by stopping for a squirrel. On the other hand, if a ball rolls out, hitting the brakes and causing a rear-end pile-up with no significant injuries might be considered a good trade-off to avoid running over a child. So the AI would have to know the difference. 67.164.113.165 (talk) 20:26, 27 July 2019 (UTC)

"There just wasn't any challenge left in it, like being the 2nd car to hit a squirrel." SinisterLefty (talk) 20:49, 27 July 2019 (UTC)

Reminds me of George Constanza and pigeons. 93.136.43.218 (talk) 22:03, 27 July 2019 (UTC)