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

AI competitions don’t produce useful models

lukeoakdenrayner.wordpress.com

Post reply on HN
Print view View original

31–40 of 64 posts

Re: AI competitions don’t produce useful models

#31

Post by Mathnerd314 »

lukeor wrote: post #24
Mathnerd314 wrote: post #22

I guess the calculator he used is https://select-statistics.co.uk/calculators/sample-size-calc... . The calculator includes this caveat: > "If one or both of the sample proportions are close to 0 or 1 then this approximation is not valid and you need to consider an alternative sample size calculation method." 0.86 is fairly close to 1. And they're not proportions but rather averages of Dice coefficients. The statisti…

Hi, author here. There are a range of ways the estimates can be improved, although many require data that isn't available. The main point is that having a ballpark idea of how reliable your results are is good, and you can achieve that with this sort of simple napkin maths. No statistician would do what I did for a formal publication, but I think what I did gets the point across.

Yes, it's a very thought-provoking article. I'm sure there are many competitions on Kaggle that were won due to the testing/training splits or other incidental choices rather than better machine learning.

But I don't think it's fair to complain about $30,000 prizes being awarded to first rather than second place in a specific competition without doing at least a little checking of whether that was actually the case. And the article kind of reads like cynicism that all machine learning is a waste, and all the algorithms are just producing random numbers that randomly happen to be right some of the time and win the competition by random chance.

Re: AI competitions don’t produce useful models

#32

Post by AstralStorm »

jph00 wrote: post #25
Mathnerd314 wrote: post #22

I guess the calculator he used is https://select-statistics.co.uk/calculators/sample-size-calc... . The calculator includes this caveat: > "If one or both of the sample proportions are close to 0 or 1 then this approximation is not valid and you need to consider an alternative sample size calculation method." 0.86 is fairly close to 1. And they're not proportions but rather averages of Dice coefficients. The statisti…

You're spot on - the calculations in the post are totally wrong. For instance, a Dice coefficient is itself the average of hundreds of thousands of observations (each pixel in an image), so you can't just treat the coefficient as a single data point. However you also can't just take a standard error, since the points are highly spatially correlated. The post also doesn't draw on any actual reported experiences of eit…

Have you actually tried to use one of those networks on a decent real dataset? They're making common category errors. ImageNet roulette is an interesting fun example of what happens if you feed an overfitted network with real images. Such as misclassifying children. :)

Re: AI competitions don’t produce useful models

#33

Post by AstralStorm »

skybrian wrote: post #21
mlnewbie wrote: post #19

Earlier quoted context omitted.

Isn't that the same thing as "not producing useful models"? Like, sure, some of the models may work, but unless you know which ones you can't make use of them.

Yes, very true, but if we're still unsure it may be worth testing them more, while if we've proven they don't work we can abandon them.

How do you test them more, with which nonbiased dataset that does not exist?

What you could do is actually describe the kind of errors the network makes. In the example of CT, false positive, false negative, wrong diagnosis. We can try to analyze what the network is detecting, rather than accept a result on some test set as real.

The millions of trials is an overstatement, but indeed few hundred thousands are needed to actually discern a winner, presuming the network did not cheat by focusing on, say, population statistics - say, certain cranium sizes being more likely to present with problems. Relying on population statistics derived from a small sample (even if representative, which it's not) is very risky...

Re: AI competitions don’t produce useful models

#34

Post by salty_biscuits »

My comment on this would be that the winner of a particular competition might not be the best model (in fact they are likely to be sneaky, e.g.extract information about the distribution of the hold out set and tune for that), but they are a great way to get a survey of the state of the art by looking in general at the types of approaches used by the top competitors. There is definitely information there.

Re: AI competitions don’t produce useful models

#35

Post by AstralStorm »

mlvsepi wrote: post #10

Sigh this beautifully captures the tragedy of biostats and epidemiology which is held hostage by lack of systems thinking. Unlike say a clinical trials, ML competitions are not limited to data available at start of the competition in fact by having a fair measurement of performance there is strong incentive to label and share more data and run checkpointed models. Further since the goal is not “publishing” with arbit…

You're making a critical mistake. Why make a network for a competition that won't produce great results without major modifications on real data?

That is the main problem, and the lack of systems thinking is on your side. There is a strong pressure to cheat and overfit. Sharing in fact makes this even stronger.

We've had some fun with that when trying to use ML for something as complex as music rhythm envelope extraction. (Which is easy in comparison to CT test.)

Best results were approaching 90% accuracy on the big suite, but real results were closer to 40%. A slightly worse solution did reliably 70%. (And was not a neural network even, plus possible to improve.)

General best approaches sometimes indeed work, but sometimes (often?) they are overfitted in architecture, not even dataset.

Re: AI competitions don’t produce useful models

#36

Post by fishooter »

The post talks about coin flipping, or 0/1 classification. Many competitions use different scores however - multiclasses, learning to find bounding boxes of objects, etc. It is much less likely to find "good" answers on the test set by chance. I think the points in the article are important, but with this context become a non-issue, when a random answer is unlikely to be correct.

Re: AI competitions don’t produce useful models

#37

Post by bermanoid »

AstralStorm wrote: post #32
jph00 wrote: post #25

Earlier quoted context omitted.

You're spot on - the calculations in the post are totally wrong. For instance, a Dice coefficient is itself the average of hundreds of thousands of observations (each pixel in an image), so you can't just treat the coefficient as a single data point. However you also can't just take a standard error, since the points are highly spatially correlated. The post also doesn't draw on any actual reported experiences of eit…

Have you actually tried to use one of those networks on a decent real dataset? They're making common category errors. ImageNet roulette is an interesting fun example of what happens if you feed an overfitted network with real images. Such as misclassifying children. :)

ImageNet Roulette deliberately uses a terrible categorization scheme that has long been acknowledged as so poor as to not admit meaningful results in order to make the highly political point that ML should never be applied to people. There's a reason most people scrub that whole piece of the taxonomy before training.

Good Resnet models trained on ImageNet (the good parts, not just people) tend to result in state of the art results for almost every transfer-learning domain they're applied to.

Re: AI competitions don’t produce useful models

#38

Post by MrMoenty »

fishooter wrote: post #36

The post talks about coin flipping, or 0/1 classification. Many competitions use different scores however - multiclasses, learning to find bounding boxes of objects, etc. It is much less likely to find "good" answers on the test set by chance. I think the points in the article are important, but with this context become a non-issue, when a random answer is unlikely to be correct.

The article is not about models being indistinguishable from random classifiers, the difference there should be very significant even on the tasks it discussed. Instead, the problem originates from the small differences in test set performance between the top N models. While that difference may very well increase when moving from binary classification to a more technically involved regression task, that is by no means guaranteed, and the main points of the article still apply.

Re: AI competitions don’t produce useful models

#39

Post by lukeor »

Mathnerd314 wrote: post #31
lukeor wrote: post #24

Earlier quoted context omitted.

Hi, author here. There are a range of ways the estimates can be improved, although many require data that isn't available. The main point is that having a ballpark idea of how reliable your results are is good, and you can achieve that with this sort of simple napkin maths. No statistician would do what I did for a formal publication, but I think what I did gets the point across.

Yes, it's a very thought-provoking article. I'm sure there are many competitions on Kaggle that were won due to the testing/training splits or other incidental choices rather than better machine learning. But I don't think it's fair to complain about $30,000 prizes being awarded to first rather than second place in a specific competition without doing at least a little checking of whether that was actually the case.…

All I can really say is that my usual readers understand that I am pro-ML, in fact I'm probably more hung go about the potential of deep learning than many of my compatriots.

I've fallen victim of getting a Twitter bump, and assuming that people know I'm not anti-ML.

The blog post is meant to be educational, not argumentative. Since it has got wider exposure I'll do a follow up to clarify my position on imagenet.

Re: AI competitions don’t produce useful models

#40

Post by skinner_ »

vaishaalshankar wrote: post #11
ajtulloch wrote: post #4

https://arxiv.org/abs/1902.10811 is an useful counterpoint to this article's comments on ImageNet overfitting (esp e.g. §3.3, "Few Changes in the Relative Order").

One of the authors of above paper here. There is a whole line of work that supports the notion that all progress we are making via this leaderboard (or competition) mechanism is in fact real progress -- and that we are not simply overfitting to the test set. This thread by Moritz Hardt does a good job laying a few reasons why this may the case: https://twitter.com/mrtz/status/1134158716251516928

I'd like to signal boost this, it's a very important line of research. Congrats on a groundbreaking paper! The first time I have seen it, I was completely shocked by that perfect linear fit on "Do ImageNet Classifiers Generalize to ImageNet?" Figure 1.

To elaborate for the other commenters: vaishaalshankar's team has created a new ImageNet evaluation dataset from scratch, and they observed that the leaderboard positions of popular image recognition models didn't change much when switching to the new evaluation. The actual performance of the models decreased significantly, but without affecting the ranking.

The OP starts from a not very controversial claim: there's a good chance that the winner of a Kaggle competition is not actually better than any of the other top k contestants, for quite large values of k. But then he completely overplays his hand, and by the time he gets to talk about ImageNet, he makes claims that were actually falsified by vaishaalshankar's paper.

Post reply on HN