68 points by bookofjoe 2 days ago | 29 comments
whatshisface 2 days ago
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?
psb217 2 days ago
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.
godelski 2 days ago
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.
godelski 2 days ago
> 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.
schmidtleonard 2 days ago
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.
godelski 1 day ago
mistermann 2 days ago
If you're the best, resting on one's laurels is not an uncommon consequence.
Componica 2 days ago
Between that time in the early 2000s I was selling implementations of really good object classifiers and OCRs.
jonas21 2 days ago
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.
Componica 2 days ago
mturmon 2 days ago
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.
Componica 1 day ago
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.
nobodyandproud 21 hours ago
One taught all of the data mining/ML algorithms including SVMs, and was clearly on their way up.
The other was relegated to teaching a couple of ANN courses and was backwatered.
The agreement was that they wouldn’t overlap in topics. Yet the first professor would take subtle couldn’t help but to take one or two swipes at ANNs when discussing SVMs.
leoc 2 days ago
fuzzfactor 2 days ago
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.