This would not happen as often if professors had time to read, instead of being under pressure to write. The only external incentive to read is so that you won't be turned down for lack of novelty, and relative to that metric a paper the field does not remember is better left nonexistent, and forgetting is as good as creating a future discovery.
In an environment where the only effect of going around popularizing information from the previous decade is interfering with other people's careers, it is no wonder that it does not happen. How did we end up with an academic system functioning as the only institution in the world with a reward for ignorance?
I figure a reasonable rule of thumb is that if someone got to the top of some system by maximizing some metric X, where X is the main metric of merit in that system, then they're unlikely to push for the system to prefer some other metric Y, even if Y is more aligned with the stated goals of the system. Pushing for a shift from X-based merit to Y-based merit would potentially imply that they're no longer sufficiently meritorious to rule the system.
To your last point, I think a lot of systems reward ignorance in one way or another. Eg, plausible denial, appearance of good intent, and all other sorts of crap that can be exploited by the unscrupulous.
While that's true, there's always exceptions. So I wouldn't say this with a defeated attitude. But I think it is also important to recognize that you can never measure things directly, it is always a proxy. Meaning that there's always a difference between what you measure and your actual goals. But I don't think it is just entrenched people that don't want to recognize this. The truth is that this means problems are much more complex and uncertain than we like. But there's actually good reason to believe the simple story, because the person selling that has clear evidence to their case while the nature of the truth is "be careful" or "maintain skepticism" is not only less exciting, it is, by nature, more abstract.
Despite this, I think getting people to recognize that measures are proxies, that they are things that must be interpreted rather than read, is a powerful force when it comes to fixing these issues. After all, even if you remove those entrenched and change the metrics, you'll later end up again with entrenchment. This isn't all bad, as time to entrenchment matters, but we should try to make that take as long as possible and try to fix things before entrenchment happens. It's much easier to maintain a clean house than to clean a dirty one. It's the small subtle things that add up and compound.
> so that you won't be turned down for lack of novelty
I think this is also a reason for lots of fraud. It can be flat out fraud, it can be subtle exaggerations because you might know or have a VERY good hunch something is true but can't prove or have the resources to prove (but will if you get this work through), or the far more common obscurification. The latter happens a lot because if something is easy to understand, it is far more likely to be seen as not novel and if communicated too well it may be even viewed as obvious or trivial. It does not matter if no one else has done it or how many people/papers you quote that claim the opposite.
On top of this, novelty scales extremely poorly. As we progress more, what is novel becomes more subtle. As we see more ideas the easier it is to relate one idea to another.
But I think the most important part is that the entire foundation of science is replication. So why do we have a system that not only does not reward the most important thing, but actively discourages it? You cannot confirm results by reading a paper (though you can invalidate by reading). You can only confirm results by repeating. But I think the secret is that you're going to almost learn something new, though information gain decreases with number of replications.
We have a very poor incentive system which in general relies upon people acting in good faith. It is a very hard system to solve but the biggest error is to not admit that it is a noisy process. Structures can only be held together by high morals when the community is small and there is clear accountability. But this doesn't hold at scale, because there are always incentives to cut corners. But if you have to beat someone who cuts corners it is much harder to do so without cutting more corners. It's a slow death, but still death.
> the entire foundation of science is replication. So why do we have a system that...
Because science is just like a software company that has outgrown "DIY QA": even as the problem becomes increasingly clear, nobody on the ground wants to be the one to split off an "adversarial" QA team because it will make their immediate circumstances significantly worse, even though it's what the company needs.
I wouldn't extrapolate all the way to death, though. If there are enough high-profile fraud busts that funding agencies start to feel political heat, they will suddenly become willing to fund QA. Until that point, I agree that nothing will happen and the problem will get steadily worse until it does.
I think I would say short term rewards heavily outweigh long term rewards. This is even true when long term rewards are much higher and even if the time to reward is not much longer than the short version. Time is important, but I think greatly over valued.
It's not the case that Bell etc. were simply overlooked: the whole question of experimental tests of interpretations of quantum mechanics was actively stigmatised and avoided by physicists until well into the '70s, at least. Clauser couldn't get a job in the area. https://arxiv.org/abs/physics/0508180
The Yann LeCun paper 'Gradient-Based Learning Applied to Document Recognition' specified the modern implementation of a convolutional neural network and was published in 1998. AlexNet, which woke up the world to CNNs, was published in 2012.
Between that time in the early 2000s I was selling implementations of really good object classifiers and OCRs.
It's not like people had been ignoring Yann LeCun's work prior to AlexNet. It received quite a few citations and was famously used by the US Postal Service for reading handwritten digits.
AlexNet happened in 2012 because the conditions necessary to scale it up to more interesting problems didn't exist until then. In particular, you needed:
- A way to easily write general-purpose code for the GPU (CUDA, 2007).
- GPUs with enough memory to hold the weights and gradients (~2010 - and even then, AlexNet was split across 2 GPUs).
- A popular benchmark that could demonstrate the magnitude of the improvement (ImageNet, 2010).
Additionally, LeCun's early work in neural networks was done
at Bell Labs in the late 80s and early 90s. It was patented by Bell Labs, and those patents expired in the late 2000s and early 2010s. I wonder if that had something to do with CNNs taking off commercially in the 2010s.
My take during that era was neural nets were considered taboo after the second AI winter of the early 90s. For example, I once proposed a start-up to consider a CNN as an alternative to their handcrafted SVM for detecting retina lesions. The CEO scoffed, telling me neural networks were dead only to acknowledge they were wrong a decade later. Younger people today might not understand, but there was a lot of pushback if you even considered using a neural network during those years. At the time, people knew that multi-layered neural networks had potential, but we couldn’t effectively train them because machines weren't fast enough, and key innovations like ReLU, better weight initializations, and optimizers like Adam didn't exist yet. I remember it taking 2-3 weeks to train a basic OCR model on a desktop pre-GPU. It wasn't until Hinton's 2006 work on Restricted Boltzmann Machines that interest in what we now call deep learning started to grow.
> My take during that era was neural nets were considered taboo after the second AI winter of the early 90s.
I'm sure there is more detail to unpack here (more than one paragraph, either yours or mine, can do). But as written this isn't accurate.
The key thing missing from "were considered taboo ..." is by whom.
My graduate studies in neural net learning rates (1990-1995) were supported by an NSF grant, part of a larger NSF push. The NeurIPS conferences, then held in Denver, were very well-attended by a pretty broad community during these years. (Nothing like now, of course - I think it maybe drew ~300 people.) A handful of major figures in the academic statistics community would be there -- Leo Breiman of course, but also Rob Tibshirani, Art Owen, Grace Wahba (e.g., https://papers.nips.cc/paper_files/paper/1998/hash/bffc98347...).
So, not taboo. And remember, many of the people in that original tight NeurIPS community (exhibit A, Leo Breiman; or Vladimir Vapnik) were visionaries with enough sophistication to be confident that there was something actually there.
But this was very research'y. The application of ANNs to real problems was not advanced, and a lot of the people trying were tinkerers who were not in touch with what little theory there was. Many of the very good reasons NNs weren't reliably performing well are (correctly) listed in your reply starting with "At the time".
If you can't reliably get decent performance out of a method that has such patchy theoretical guidance, you'll have to look elsewhere to solve your problem. But that's not taboo, that's just pragmatic engineering consensus.
You're probably right in terms of the NN research world, but I've been staring at a wall reminiscing for a 1/2 hour and concluded... Neural networks weren’t widely used in the late 90s and early 00s in the field of computer vision.
Face detection was dominated by Viola-Jones and Haar features, facial feature detection relied on active shape and active appearance models (AAMs), with those iconic Delaunay triangles becoming the emblem of facial recognition. SVMs were used to highlight tumors, while kNNs and hand-tuned feature detectors handled tumors and lesions. Dynamic programming was used to outline CTs and MRIs of hearts, airways, and other structures, Hough transforms were used for pupil tracking, HOG features were popular for face, car, and body detectors, and Gaussian models & Hidden Markov Models were standard in speech recognition. I remember seeing a few papers attempting to stick a 3-layer NN on the outputs of AAMs with limited success.
The Yann LeCun paper felt like a breakthrough to me. It seemed biologically plausible, given what I knew of the Neocognitron and the visual cortex, and the shared weights of the kernels provided a way to build deep models beyond one or two hidden layers.
At the time, I felt like Cassandra, going from past colleagues and computer vision-based companies in the region, trying to convey to them just how much of a game changer that paper was.
I wish I had access to papers for free, I'd read more and do more things.
For example, the earliest magnetic gears papers were $25 each and I needed about ten that cited each other.
That's why I didn't try to create a magnetic hub for cycling.
Att the time I thought I could make a more compact geared hub, but needed the torque calculations to be sure.
I was a college student, my university did not have access to those journals, and I had no money.
Interesting that some communities (apparently physics) use "arXiv" with the definite article ("the arXiv"), but in machine learning / CS we always say simply "arXiv". I went and checked, and the official site doesn't use an article (https://info.arxiv.org/about/index.html)
This is fascinating and I think is one of the major areas that new AI systems will impact humanity, ie by combing through millions of papers to make connections and discover such sleeping beauties.
BTW, I noticed a similar phenomenon on HN submissions (on much shorter timescales): sometimes they just for a few hours with 2-3 points and then shoot up.
I think studying this stuff is always going to seem mysterious unless you account for the concept of fashion in science. Specifically what I mean is that two papers (or ideas or approaches) X and Y can have equal “objective scientific merit” but X is more popular than Y because of random initial conditions (e.g., a famous researcher happened upon X first and started mentioning it in their talks) that are self-reinforcing. The root cause of this phenomenon is that most/all researchers can’t justify what they work on from first principles; for both good and bad reasons, they ultimately rely on the wisdom of the crowd to make choices about what to study and cite. This naturally leads to big “flips” when a critical mass of people realize that Y is better than X, and then suddenly everyone switches en mass.
An interesting example of someone who managed to produce two unrelated "sleeping beauties" in different fields is the German mathematician Grete Hermann. Her Ph.D. thesis in the 20s gave effective algorithms for many questions in abstract algebra. I think her motivation was philosophical, that an effective algorithm is better than an abstract existence result, but it wasn't considered that interesting of a question until computers were invented and computer algebra developed, and then immediately several of her algorithms became the state-of-the-art at the time.
Unrelatedly, she wrote on the foundations of quantum mechanics, and showed that a "theorem" of John von Neumann, which would have ruled out later research by Bohm and Bell if it were correct, was false three years after he published it. Bohm and Bell had to independently rediscover that the result was false years later.
I’m skeptical of LLM’s ability to reason, but trawling through the vast research literature is an area where they can shine. They can both summarize and serve as a superior search engine.
This would not happen as often if professors had time to read, instead of being under pressure to write. The only external incentive to read is so that you won't be turned down for lack of novelty, and relative to that metric a paper the field does not remember is better left nonexistent, and forgetting is as good as creating a future discovery.
In an environment where the only effect of going around popularizing information from the previous decade is interfering with other people's careers, it is no wonder that it does not happen. How did we end up with an academic system functioning as the only institution in the world with a reward for ignorance?
I figure a reasonable rule of thumb is that if someone got to the top of some system by maximizing some metric X, where X is the main metric of merit in that system, then they're unlikely to push for the system to prefer some other metric Y, even if Y is more aligned with the stated goals of the system. Pushing for a shift from X-based merit to Y-based merit would potentially imply that they're no longer sufficiently meritorious to rule the system.
To your last point, I think a lot of systems reward ignorance in one way or another. Eg, plausible denial, appearance of good intent, and all other sorts of crap that can be exploited by the unscrupulous.
While that's true, there's always exceptions. So I wouldn't say this with a defeated attitude. But I think it is also important to recognize that you can never measure things directly, it is always a proxy. Meaning that there's always a difference between what you measure and your actual goals. But I don't think it is just entrenched people that don't want to recognize this. The truth is that this means problems are much more complex and uncertain than we like. But there's actually good reason to believe the simple story, because the person selling that has clear evidence to their case while the nature of the truth is "be careful" or "maintain skepticism" is not only less exciting, it is, by nature, more abstract.
Despite this, I think getting people to recognize that measures are proxies, that they are things that must be interpreted rather than read, is a powerful force when it comes to fixing these issues. After all, even if you remove those entrenched and change the metrics, you'll later end up again with entrenchment. This isn't all bad, as time to entrenchment matters, but we should try to make that take as long as possible and try to fix things before entrenchment happens. It's much easier to maintain a clean house than to clean a dirty one. It's the small subtle things that add up and compound.
On top of this, novelty scales extremely poorly. As we progress more, what is novel becomes more subtle. As we see more ideas the easier it is to relate one idea to another.
But I think the most important part is that the entire foundation of science is replication. So why do we have a system that not only does not reward the most important thing, but actively discourages it? You cannot confirm results by reading a paper (though you can invalidate by reading). You can only confirm results by repeating. But I think the secret is that you're going to almost learn something new, though information gain decreases with number of replications.
We have a very poor incentive system which in general relies upon people acting in good faith. It is a very hard system to solve but the biggest error is to not admit that it is a noisy process. Structures can only be held together by high morals when the community is small and there is clear accountability. But this doesn't hold at scale, because there are always incentives to cut corners. But if you have to beat someone who cuts corners it is much harder to do so without cutting more corners. It's a slow death, but still death.
> the entire foundation of science is replication. So why do we have a system that...
Because science is just like a software company that has outgrown "DIY QA": even as the problem becomes increasingly clear, nobody on the ground wants to be the one to split off an "adversarial" QA team because it will make their immediate circumstances significantly worse, even though it's what the company needs.
I wouldn't extrapolate all the way to death, though. If there are enough high-profile fraud busts that funding agencies start to feel political heat, they will suddenly become willing to fund QA. Until that point, I agree that nothing will happen and the problem will get steadily worse until it does.
I think I would say short term rewards heavily outweigh long term rewards. This is even true when long term rewards are much higher and even if the time to reward is not much longer than the short version. Time is important, but I think greatly over valued.
World class guerilla marketing might have something to do with it, it is arguably the most adored institution in existence.
If you're the best, resting on one's laurels is not an uncommon consequence.
It's not the case that Bell etc. were simply overlooked: the whole question of experimental tests of interpretations of quantum mechanics was actively stigmatised and avoided by physicists until well into the '70s, at least. Clauser couldn't get a job in the area. https://arxiv.org/abs/physics/0508180
The Yann LeCun paper 'Gradient-Based Learning Applied to Document Recognition' specified the modern implementation of a convolutional neural network and was published in 1998. AlexNet, which woke up the world to CNNs, was published in 2012.
Between that time in the early 2000s I was selling implementations of really good object classifiers and OCRs.
It's not like people had been ignoring Yann LeCun's work prior to AlexNet. It received quite a few citations and was famously used by the US Postal Service for reading handwritten digits.
AlexNet happened in 2012 because the conditions necessary to scale it up to more interesting problems didn't exist until then. In particular, you needed:
- A way to easily write general-purpose code for the GPU (CUDA, 2007).
- GPUs with enough memory to hold the weights and gradients (~2010 - and even then, AlexNet was split across 2 GPUs).
- A popular benchmark that could demonstrate the magnitude of the improvement (ImageNet, 2010).
Additionally, LeCun's early work in neural networks was done at Bell Labs in the late 80s and early 90s. It was patented by Bell Labs, and those patents expired in the late 2000s and early 2010s. I wonder if that had something to do with CNNs taking off commercially in the 2010s.
My take during that era was neural nets were considered taboo after the second AI winter of the early 90s. For example, I once proposed a start-up to consider a CNN as an alternative to their handcrafted SVM for detecting retina lesions. The CEO scoffed, telling me neural networks were dead only to acknowledge they were wrong a decade later. Younger people today might not understand, but there was a lot of pushback if you even considered using a neural network during those years. At the time, people knew that multi-layered neural networks had potential, but we couldn’t effectively train them because machines weren't fast enough, and key innovations like ReLU, better weight initializations, and optimizers like Adam didn't exist yet. I remember it taking 2-3 weeks to train a basic OCR model on a desktop pre-GPU. It wasn't until Hinton's 2006 work on Restricted Boltzmann Machines that interest in what we now call deep learning started to grow.
> My take during that era was neural nets were considered taboo after the second AI winter of the early 90s.
I'm sure there is more detail to unpack here (more than one paragraph, either yours or mine, can do). But as written this isn't accurate.
The key thing missing from "were considered taboo ..." is by whom.
My graduate studies in neural net learning rates (1990-1995) were supported by an NSF grant, part of a larger NSF push. The NeurIPS conferences, then held in Denver, were very well-attended by a pretty broad community during these years. (Nothing like now, of course - I think it maybe drew ~300 people.) A handful of major figures in the academic statistics community would be there -- Leo Breiman of course, but also Rob Tibshirani, Art Owen, Grace Wahba (e.g., https://papers.nips.cc/paper_files/paper/1998/hash/bffc98347...).
So, not taboo. And remember, many of the people in that original tight NeurIPS community (exhibit A, Leo Breiman; or Vladimir Vapnik) were visionaries with enough sophistication to be confident that there was something actually there.
But this was very research'y. The application of ANNs to real problems was not advanced, and a lot of the people trying were tinkerers who were not in touch with what little theory there was. Many of the very good reasons NNs weren't reliably performing well are (correctly) listed in your reply starting with "At the time".
If you can't reliably get decent performance out of a method that has such patchy theoretical guidance, you'll have to look elsewhere to solve your problem. But that's not taboo, that's just pragmatic engineering consensus.
You're probably right in terms of the NN research world, but I've been staring at a wall reminiscing for a 1/2 hour and concluded... Neural networks weren’t widely used in the late 90s and early 00s in the field of computer vision.
Face detection was dominated by Viola-Jones and Haar features, facial feature detection relied on active shape and active appearance models (AAMs), with those iconic Delaunay triangles becoming the emblem of facial recognition. SVMs were used to highlight tumors, while kNNs and hand-tuned feature detectors handled tumors and lesions. Dynamic programming was used to outline CTs and MRIs of hearts, airways, and other structures, Hough transforms were used for pupil tracking, HOG features were popular for face, car, and body detectors, and Gaussian models & Hidden Markov Models were standard in speech recognition. I remember seeing a few papers attempting to stick a 3-layer NN on the outputs of AAMs with limited success.
The Yann LeCun paper felt like a breakthrough to me. It seemed biologically plausible, given what I knew of the Neocognitron and the visual cortex, and the shared weights of the kernels provided a way to build deep models beyond one or two hidden layers.
At the time, I felt like Cassandra, going from past colleagues and computer vision-based companies in the region, trying to convey to them just how much of a game changer that paper was.
Tip, meet iceberg.
Science is like music, most of it is never recorded to begin with.
Much less achieves widespread popularity.
When you restrict it to academic journals the real treasure-trove can not even be partially contained in the vessel which you are searching within.
I wish I had access to papers for free, I'd read more and do more things.
For example, the earliest magnetic gears papers were $25 each and I needed about ten that cited each other. That's why I didn't try to create a magnetic hub for cycling. Att the time I thought I could make a more compact geared hub, but needed the torque calculations to be sure. I was a college student, my university did not have access to those journals, and I had no money.
You do, they're on SciHub.
Also, ~every physics paper since 1993 is on the arXiv. The same is true for math and CS with later cutoffs.
Interesting that some communities (apparently physics) use "arXiv" with the definite article ("the arXiv"), but in machine learning / CS we always say simply "arXiv". I went and checked, and the official site doesn't use an article (https://info.arxiv.org/about/index.html)
This is fascinating and I think is one of the major areas that new AI systems will impact humanity, ie by combing through millions of papers to make connections and discover such sleeping beauties.
BTW, I noticed a similar phenomenon on HN submissions (on much shorter timescales): sometimes they just for a few hours with 2-3 points and then shoot up.
I think studying this stuff is always going to seem mysterious unless you account for the concept of fashion in science. Specifically what I mean is that two papers (or ideas or approaches) X and Y can have equal “objective scientific merit” but X is more popular than Y because of random initial conditions (e.g., a famous researcher happened upon X first and started mentioning it in their talks) that are self-reinforcing. The root cause of this phenomenon is that most/all researchers can’t justify what they work on from first principles; for both good and bad reasons, they ultimately rely on the wisdom of the crowd to make choices about what to study and cite. This naturally leads to big “flips” when a critical mass of people realize that Y is better than X, and then suddenly everyone switches en mass.
Granted. I'm still trying to find out what led up to several people revisiting Mendel's work in 1900. https://en.wikipedia.org/wiki/Mendelian_inheritance
Related. Others?
The World Is Full of Sleeping Beauties - https://news.ycombinator.com/item?id=35975866 - May 2023 (1 comment)
An interesting example of someone who managed to produce two unrelated "sleeping beauties" in different fields is the German mathematician Grete Hermann. Her Ph.D. thesis in the 20s gave effective algorithms for many questions in abstract algebra. I think her motivation was philosophical, that an effective algorithm is better than an abstract existence result, but it wasn't considered that interesting of a question until computers were invented and computer algebra developed, and then immediately several of her algorithms became the state-of-the-art at the time.
Unrelatedly, she wrote on the foundations of quantum mechanics, and showed that a "theorem" of John von Neumann, which would have ruled out later research by Bohm and Bell if it were correct, was false three years after he published it. Bohm and Bell had to independently rediscover that the result was false years later.
I’m skeptical of LLM’s ability to reason, but trawling through the vast research literature is an area where they can shine. They can both summarize and serve as a superior search engine.
Just looking at the headline, I was expecting to see a few 10s after the link.