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

Alphafold

github.com

Post reply on HN
Print view View original

141–150 of 170 posts

Re: Alphafold

#141

Post by boublepop »

evouga wrote: post #124
sbierwagen wrote: post #96

Earlier quoted context omitted.

>A disturbing thing is that the architecture is much less novel than I originally thought it would be, so this shows perhaps one of the major difficulties was having the resources to try different things on a massive set of multiple alignments. A similar concern has sparked some worries about "AI overhang" https://www.lesswrong.com/posts/75dnjiD8kv2khe9eQ/measuring-... Most of the compute in ML research seems to be g…

Self-driving is a cursed problem, but the thorniest obstacles relate to optics and politics and not technology. AI can already replace some human drivers some of the time; but it doesn’t matter because each news story about a Tesla killing a passenger or driving into a parked fire truck sets back public acceptance of self-driving cars by years.

True, and Tesla is so far behind competitors when it comes to self driving that they knowingly push the envelope, because either they succeed, or they kill their customers but hurt all self-driving companies. So both scenarios work to their advantage and they just pay the minor fines and compensations to the people the kill in the process. And the Elon posts some Tweets blaming the people who got killed and his fandome happily cheers.

There is a reason Waymo is progressing at what looks like snail speed from outside observers, and that reason is that it’s the only ethical pace for development and testing of what could eventually become broadly available fully autonomous vehicles.

Re: Alphafold

#142

Post by Jyaif »

TaupeRanger wrote: post #84
timr wrote: post #23

Earlier quoted context omitted.

> A disturbing thing is that the architecture is much less novel than I originally thought it would be, so this shows perhaps one of the major difficulties was having the resources to try different things on a massive set of multiple alignments. This is something an industrial lab like DeepMind excels at. Whereas universities tend to suck at anything that requires a directed effort of more than a handful of people. Y…

That's the case with basically everything DeepMind does. They have a very good PR department which hypes up everything they do while conveniently ignoring that basically nothing of any practical consequence has come of their endeavors. But I do think it's important that these companies exist now so we can see what not to try going forward.

Their PR strategy is to take problems people thought were impossible to solve in the next 10 years, and solve them (Go) or nearly solve them (StarCraft 2, protein solving)

Re: Alphafold

#143

Post by allyourhorses »

tdfirth wrote: post #129

This isn't a criticism - I'm just curious to hear people's thoughts on this. When I look at this code, one of my initial reactions is that it does not seem to be very thoroughly tested. Sure, certain modules have been tested (e.g. `model.quat_affine`) but it's not clear how completely. Meanwhile, other modules, for example `model.folding`, have not been tested at all, despite containing large amounts of complex logic…

Prior to this model, protein folding hadn't seen significant advancements in a decade or more. Worrying about the lack of tests in a first of its kind model is very much akin to complaining about the choice of font in a user manual for the world's first warp drive. I understand you're attempting to frame the problem in terms in things you know, but trying to weigh down pioneering research with professional development ceremony is very much counterproductive. The 'missing' ceremony would not have contributed to the strength of AlphaFold's result, the model's only purpose was to compete within the context of an existing validation framework.

Re: Alphafold

#144

Post by pelorat »

boublepop wrote: post #141
evouga wrote: post #124

Earlier quoted context omitted.

Self-driving is a cursed problem, but the thorniest obstacles relate to optics and politics and not technology. AI can already replace some human drivers some of the time; but it doesn’t matter because each news story about a Tesla killing a passenger or driving into a parked fire truck sets back public acceptance of self-driving cars by years.

True, and Tesla is so far behind competitors when it comes to self driving that they knowingly push the envelope, because either they succeed, or they kill their customers but hurt all self-driving companies. So both scenarios work to their advantage and they just pay the minor fines and compensations to the people the kill in the process. And the Elon posts some Tweets blaming the people who got killed and his fando…

Teslas philosophy is that self driving is useless if it can't handle every road and route out there, while Waymo who already has a ride hailing service, relies on pre-mapped routes and databases.

Both have their advantages and disadvantages. Waymo requires huge databases and a constant network connection. But it's good enough to be used in real-life without a backup driver, in certain select locations that is. If there's a road construction somewhere, don't expect the Waymo car to handle it unless they update their database.

Tesla on the other hand attempts to solve the problem with minimal database use, where most of the info coming from road markings and traffic signs. A much harder problem.

Re: Alphafold

#145

Post by nl »

neolog wrote: post #117
nl wrote: post #110

Earlier quoted context omitted.

> Most of the compute in ML research seems to be going into architecture search. No it's not. Only Google spends significant time with automatic architecture search, and many people think this is really to try to sell cloud capacity. > Once the architecture is found, training and net finetuning/transfer learning is comparatively cheap Training isn't cheap for significant problems. Getting the data is very expensive,…

parent's EfficientNet graph seems really dramatic. Is it misleading somehow?

It's not misleading assuming you know the field.

If you don't you might not realize they are comparing ResNet (designed for ultimate performance) vs EffcientNet to show how close it gets in accuracy in the FLOPs budget.

Note that the best accuracy for EffcientNet is roughly 1-2% better than ResNet/ResNeXT/SENet/etc but does have a much better FLOPs budget.

But these other architectures were never optimised for FLOPs. It wasn't event a consideration when designing them.

And EffcientNet is about a (manually designed) technique for scaling neural networks up in accuracy. Only EffcientNet-B0 is designed by AutoML, the others are scaled up. See the paper[1] for complete details.

Like-for-like should be against MobileNet etc. EfficentNet is better, but the comparison is more reasonable in general.

https://machinethink.net/blog/mobile-architectures/ is a really good overview.

[1] https://arxiv.org/pdf/1905.11946.pdf

Re: Alphafold

#146

Post by dekhn »

exikyut wrote: post #119
dekhn wrote: post #95

Earlier quoted context omitted.

there are research groups this would be useful for but structures are not on the critical path to drug discovery or approval.

Out of (probably overoptimistic :) ) curiosity, what do you see are the critical paths?

Initial molecule generation and FDA approval.

Re: Alphafold

#147

Post by dekhn »

ehsankia wrote: post #118
devindotcom wrote: post #76

Also announced today was RoseTTAFold from UW's Baker Lab, which claims nearly the same accuracy at much higher efficiencies. There's a public server and paper in Science. More info here and here: https://www.bakerlab.org/index.php/2021/07/15/accurate-prote... https://techcrunch.com/2021/07/15/researchers-match-deepmind...

Could it be that AlphaFold 2 was open sourced in response to this?

it's very likely the baker submission to science forced DM's hand.

Re: Alphafold

#148

Post by dekhn »

Jyaif wrote: post #142
TaupeRanger wrote: post #84

Earlier quoted context omitted.

That's the case with basically everything DeepMind does. They have a very good PR department which hypes up everything they do while conveniently ignoring that basically nothing of any practical consequence has come of their endeavors. But I do think it's important that these companies exist now so we can see what not to try going forward.

Their PR strategy is to take problems people thought were impossible to solve in the next 10 years, and solve them (Go) or nearly solve them (StarCraft 2, protein solving)

DM didn't "solve" anything around proteins. They just made an improvement to existing homology modelling methods. If you look, the system is incredibly dependent on having large numbers of high quality sequence alignments to proteins with known structure and lots and lots of evolutionary data.

This was actually fairly obvious 20 years ago and it's been disappointing to me to see how long it took somebody to make this improvement, but really, it couldn't have been done without recent algorithmic improvements and huge amounts of CPU time.

Re: Alphafold

#149

Post by haihaibye »

Ultimatt wrote: post #61
Cas9 wrote: post #38

Honest question: since AlphaFold doesn't really _solve_ the protein folding problem (it's NP-complete after all), but only _approximates_ solutions very well, what are the real impacts of this? Isn't a good approximation of a protein enough to cause unexpected problems? How do we know that an approximate structure will perform the same as the correct solution?

There is a lot of bias in the chat here from a more chemistry and pharma slant. If you ignore this AlphaFold solves in a very meaningful way the problem blocking a lot of science investigation. For comparative and evolutionary analysis structure is far more conserved than sequence. Especially in things like viruses or anything with a high rate of reproduction like bacteria. Just knowing the general fold or overall st…

> even if their genomic sequence is completely dissimilar

I think you mean amino acid homology? (due to synonymous mutations)

I looked it up and you're right, protein structure/motifs are much more highly conserved than amino acid sequence https://humgenomics.biomedcentral.com/articles/10.1186/1479-...

Re: Alphafold

#150

Post by dekhn »

dmos62 wrote: post #140
benschulz wrote: post #130

Earlier quoted context omitted.

My understanding is that it has been manually tested. I.e. it has produced correct results to previously intractable problems. I'm not sure how much automated testing would add at that point.

Unit testing usually isn't easily replaced by manual testing. If you have, for example, 3 units that can be in 2 different modes each, that's 2^3 different combinations, but only 2*3 unit modes. Testing the end result is more work than testing the units.

Discovery science is different from web software engineering. Most discovery scientists use manual testing, not unit testing. Very few actually do integration tests or system tests (this is something I'm trying to change).

And, given the external results of the application, it's unclear to me how much additional value would come from a rigorous testing system.

Post reply on HN