Maybe I have a wrong conception of what research is supposed to achieve, but commercializing new insights is absolutely one of the intended outcomes. One would sure hope that taxpayer money isn't funneled into research to... just enable more research. At some point the public should tangibly benefit from it, which is not achieved by writing more papers.

This all notwithstanding the fact that DeepMind intends to make AlphaFold open source and available to the community.

This is a recent idea, dating back to the late 70s and implemented through the 1980 Bayh-Dole Act. Before that research was research; development (and private research of course) was the province of business.

The Gradgrind mentality that all research must be commercialized has impoverished basic research; of what use is looking for gravity waves or new branches of mathematics (just look at the contortions university press offices go to to justify some new paper on quantum mechanics).

Speaking of which, QM is a perfect example of something that would have advanced very slowly had this attitude existed 150 years ago…yet it is at the heart of the semiconductor revolution!

If so, I don't think that really applies to what the article is talking about. The article is talking about Alphabet potentially using large amounts of data from other researchers, mostly academic, who were funded by the government and commercializing it. That's more akin to how it was before Bayh-Dole: a private company taking government funded research they were not involved in, adding their own privately funded research, and making something commercial.

This means, for example, that David Baker licenses Rosetta for free to academic and government, but commercial users have to obtain a paid commercial license. Baker (his lab, or LLC, or whomever vends Rosetta Commercial) benefits monetarialy from all the data that Rosetta includes, which is decades of structural biology funded by NIH and others.

My comment was in reply to this comment by MauranKilom:

> > Maybe I have a wrong conception of what research is supposed to achieve, but commercializing new insights is absolutely one of the intended outcomes. One would sure hope that taxpayer money isn't funneled into research to... just enable more research.

And further on your comment:

> The article is talking about Alphabet [commercializing results from public datasets without needing to pay for them] That's more akin to how it was before Bayh-Dole:

Indeed, pre Bayh-Dole, publicly funded research was public (consider it public domain, or at least "MIT licensed") and anyone could use it.

Now everything has to be licensed from university licensing departments, typically with an expensive exclusive. Which has had a distorting effect on research, not merely restricting use (have you ever tried to work with a university licensing office? They consider even the most trivial results to be Nobel prize class) but, because they are a source of revenue, bending resource allocation, tenure, etc much as sports teams do for the schools that have them.

I don't think federal funding of research is that much older in the US, only really starting in the 50s apart from military research. How exactly were the early QM researchers funded anyway? (apart from Einstein's famous day job at the patent office). I know at least a few of them had fellowships at universities, meaning rich benefactors.

US government support for university research dates back to patent holder Abraham Lincoln who even in the middle of a war got legislation passed to support land grant (mostly ag) colleges and universities (and of which MIT was one of the very early beneficiaries). However it was small and you are right that in WWII the model of the US modern research university was explicitly created by James Conant, with MIT again being the largest beneficiary (note that all tuition and student expenses are about 14% of MIT's revenue and 16% of expenditures, and the number of staff is greater than that of the student body -- it's a huge government research lab with a small school attached).

The problem with this model is that unless you are MIT (/Stanford/Harvard/Cornell/CMU et al -- maybe 25 institutions, if that) licensing revenue matters, and affects who gets tenure, departmental budgets etc.

> How exactly were the early QM researchers funded anyway? (apart from Einstein's famous day job at the patent office). I know at least a few of them had fellowships at universities, meaning rich benefactors.

In Europe, in the 20th century funding came primarily from governments (and benefactors, more early in the century), under varying institutions (the big research institutions in Imperial and post-WWI Germany, "Institutes" in France, Oxbridge in the UK, etc). In The USA it was the institutions themselves, some benefactors and, as I said, some government funding (like Fermi and Lawrence).

Everything we do is built on top of what came before.

For example, research is paid for by the public, but then the products that affect people are completely captured by monopolists and spooned out in such a way to make sure only the monied sections of the population get them until the public protests enough to create a program like medicaid.

If we paid most of the cost, we should get most of the benefit. The monopolists should be happy to make any money at all, not their superprofits. Fair right?

I agree with your sentiment, but I also think it's worth thinking carefully about two of the main points that stuck out to me:

- Access to compute for large models

- Access to large datasets (in this case mostly taxpayer funded academic research)

Every company and/or research group has access to the data, but some have a huge advantage in terms of compute. If there's a question about commercializing research, the scales are tilted toward those with more compute.

In this specific case, I think the intention to make AlphaFold open source and available to the community is obviously the best solution. But my question is, what happens if a less altruistic for-profit entity uses its huge compute advantage to develop new techniques and insights, and then patents everything before it becomes available to the community?

I understand that is the basic mechanism for how medical/pharmaceutical research gets translated into life-saving treatments, but if we're approaching a generalized model that can pump out "patent-worthy" discoveries only bound by the amount of data and access to compute, there's an obvious opportunity for a winner-take-most scenario.

While that kind of semi symbiotic relationship can (and has been observed to) exist, it does so best in an an environment that looks different than what is described here (few large near monopolies, legislative regulations that are best navigated using wealth, a market that has inelastic bargaining qualities).

Perhaps with some interest, since the grants are high risk (many grants fail)

Or we could do like before: let the church help the poors, the nobles take decisions and the peasants make the food. Like that, everyone has its clear and simple role and you wont complain of taxes: you ll have no revenue :)

So long as any profits from these investments are not ploughed back into general revenues, your concerns are moot. For example, you could establish an independent body—let's call it INVEST1—to oversee these investments. INVEST1 would be required to divest ownership of successful enterprises at a threshold that ensures money spent roughly matches money earned (and thus self-funding). Once it reaches a stable equilibrium, you spin off INVEST1 as a fully independent not-for-profit. Government then establishes INVEST2 and the process starts again from the beginning. Rinse and repeat.

So when people build off humanity's past knowledge and they pay for the privilege I assume the new knowledge that is created does not belong to humanity any longer and belongs to individuals?

https://sbir.nih.gov/about/what-is-sbir-sttr

neither were at play here but the idea is pretty darn normal.

The money will be made by private companies that have no connection to the researchers who received the grants, but simply use the published research in something they build.

It's hard to see a good way to build a system to make the private companies pay back the grants. It would be an accounting and tracking nightmare to try to figure out how much money is actually being made from the research that any given grant paid for.

There’s a reason for the saying ‘standing on the shoulders of giants’

Yes indeed. The public. Not capital, not private concerns, but the public.

I think there's a legitimate argument taxes exist for this sort of thing, but (1) taxes arguably are avoided in various ways to the point it's a currently broken system, and (2) this is a rare case where the government has a clear case for a specific amount of money owed by a specific company — why not keep it simple?

If the grants aren't worth paying back at zero percent, the corporation shouldn't be taking them.

The public absolutely should have some ROI, and in fact does in the form of taxes.

I say this as an academic, of course.

In any case, left to their own devices, corporate R&D teams wouldn’t be able to define goals that work for their business. Like without the competition and goals being defined for them, DeepMind would be having meetings with brand managers about the avant- grade of ad tracking.

Getting an education (even an advanced one) is a completely separate thing than entering academia and I suspect you know this.

This is the same energy as saying that being the son of an incredibly rich person is a completely separate thing than being rich.

Private companies are much more efficient in reaching a well-defined goal.

Academia is much more efficient in reaching ill-defined goals.

The thing is that the majority of goals for basic science are very ill-defined and virtually all breakthroughs are serendipitous (reaching from antibiotics to more recently CRISPR-Cas). So I don't think it makes sense to advocate for one vs the other.

Science is pretty difficult frankly, especially these days. Sure science is worth doing, and can be quite rewarding, no doubt. But I can't quite understand the popularity. Writing papers is particularly painful, again I can't understand why there are so many papers being written, it is torrential. As others have noted, this is not the same as actual progress. But lately I have noticed something telling about several leading researchers in my field, people who have made it in every sense of the word, work at the top institutions, regularly publish stunning research in Science, Nature. They have been moving to the pharmaceutical industry in significant numbers. This tells you, basically, that academia sucks for everyone. This is also my impression from talking to people who have entered research in the last 15 years. The older researchers seem happier. Perhpas it is the pay scales that scale with seniority, or just survivorship bias.

Any bit of research into this (and most of their successes in other fields) will show otherwise.

Or you could supply references and/or an argument. The "if you researched this you'd agree with my claims" approach is pretty pernicious.

Everyone knows that Deepmind hires experts in deep learning, as well as subject matter experts. The idea that their success is just because of compute power alone is preposterous.

Even the article that this comment thread is about has extensive speculation about the techniques Deepmind used. There’s clearly a high level of anticipation to read their paper.

Concentrating all these prolific researchers in one place, removing the publish-or-perish incentive, and giving them access to unlimited data and computing power.

Seems like that could make a difference.

[1] Standard caveats apply and the point stands. [2] Standard caveats apply and the point stands.

It includes some commentary on this discovery, as well.

Is there an update since then? Have they actually done anything useful with it?

> As of 18 June 2021, according to DeepMind's CEO Demis Hassabis a full methods paper to describe AlphaFold 2 had been written up and was undergoing peer review prior to publication, which would be accompanied by open source code and "broad free access to AlphaFold for the scientific community"

The original version on github can only compute the specific systems in the paper.

https://github.com/deepmind/deepmind-research/tree/master/al...

I don't know why scientific publications keep doing PR work for them.

> The details of how AlphaFold 2 works are still unknown, and we may not have full access to them until their paper is peer-reviewed (which may take more than a year, based on their CASP13 paper).

So it's not particularly surprising that we haven't heard much yet.

Ie. Given a structure that you'd like a protein to have, develop a sequence for it.

If we could do that easily, we could start making molecular machines for all kinds of tasks. Rather than co-opting enzymes from nature, we could design our own.

So many industries could benefit from that, even if you exclude all the biomedical applications where such an approach might be considered too high risk. We could for example begin with dishwashing tablets which actually get burnt on stuff off...

If you wanted to make a more reasonable length protein, of say 100-200, you would run out of atoms in the universe to do computations with.

Or is it just such a huge model it basically encodes an entire database after weeks and weeks of computation and has a more compressed form?

The architecture of the system and design of the training methodology are laid out to specifically prevent direct database-esque "pattern in pattern out" failure mode.

Similar to Google deep dream, there will be contextual features and feature clusters encoded into neurons that can be explored and extracted, and those could provide insights that can be sourcing translated into "hard" science, with explicit formulae and theory allowing a fully transparent model to be created.

Like other transformer models, you can elicit the training data intact, but such scenarios are a statistically insignificant selection of the range of outputs the models are capable of producing. That doesn't mean anything with regards to accuracy of the novel output, though.

With alphafold 2 going open source, it's possible that tools and methodologies to extract hard science from transformers will be formalized quickly and in the public eye. We have an amazingly powerful new tool, and the coming decades will be fascinating.

- The thing I'd be curious about is whether or not "not formalized" would imply "not consistently generalizing", whether it would have to be train all over if given a problem similar to but identical with, the problem it solves.

The tools to do that with transformers in particular and neural networks in general are still pretty new and specific to particular models.

https://ai.googleblog.com/2015/06/inceptionism-going-deeper-...

Someone with extensive knowledge of protein folding and chemistry and any relevant domain would need to manually categorize different neurons, their contexts, and effects, then experiment with the model in a static/procedural inference mode, then parse out the overall functions. By reproducing well known simple instances and tracing the propagation of information through the network, you can validate what a model knows by mapping to current knowledge.

Using the model in this way can also lead to interesting real world experiment choices, by processing targets suggested by real life phenomena and then coordinating from known functions of the model and real life experiment. Different proteins in coronavirus, for example, would likely have a different cascade of activations in the model than the flu virus. Any shared features between the two might be captured in the activation patterns, possibly leading to new information about the diseases that would opaque to science without these models.

So an extremely large DL network would have a good chance to find and integrate ("compress") all the existing folds and sub-folds that human researchers or Rosetta missed or just hadn't the time to investigate and characterize yet (I'm not an expert on Rosetta by far btw so please expand if you are :).

I would venture to say it's a good problem fit for DL methods (as was impressively demonstrated).

Regarding your question, "does it understand something about the structure of the protein folding problem" - expanding on the above, I would say it understands enough, but it probably doesn't understand the generics of chemistry as protein and their folding is a biased subset. The output is (as far as I remember) an atom distance matrix and not atom trajectories etc. so folding dynamics is not part of the model (this is btw an important part of protein science as well).

From what I understand, deep learning, although opaque and relying on tons of data, is a bit magical : although one would say "it's just probabilities", it actually does probabilities at a level where it actually figures some things.

Plus, and that's very much a problem to me, Google does it at 100x the scale of a regular researcher. Since I just invested a year in studying data sciences, that worries me a lot : where am I suppose to work if, to produce meaningful results, you need way-too-expensive hardware...

If you are in a team that is looking at problems that justify massive hardware (in the sense that solving them will pay back the capital and environmental cost) then you will have access to said hardware.

Most (almost all) AI and Data Science teams are not working on that kind of problem though, and it's often the case that we are working on cloud infrastructures where GPU's and TPU's can be accessed on demand and $100's can be used to train pretty good models. Obviously models need to be trained many times so the crossover point from a few $100 to a few $1000 can be painful - but actually many/most engagements really only need models that cost <$100 to train.

Also many of the interesting problems that are out there can utilize transfer learning over shared large pretrained models such as Resnet or GPT-2 (I know that in the dizzying paced modern world these no longer count as large or modern but they are examples...) So for images and natural language problems we can get round the intractable demand for staggeringly expensive compute.

Imagine that you had got a degree in Aeronautical Engineering, you are watching the Apollo program and wondering how you will get a job at NASA or something similar... but there are lots of jobs at Boeing and Lockheed and Cessna and so on.

I know how you feel, but I also think stories like this may be a wake-up call for some groups to invest more in hardware. The Baker and Tsinghua University groups are not small. They can afford more than 4 GPUs.

Probably it's more about setting up a good pipeline. Once you get about 4 GPUs you need more than one machine to run them. Hopefully in the next years we'll see more open source tools to make it easy to "build your own GPU cloud".

Anyway these days I see a lot of industrial investment in middle size computing datacenters full of GPUs. Sure Google scale is not within reach but I'm sure there's room for scrapier algorithms and training methods to demonstrate feasibility and then paying a lump sum to AWS (or some ml-expert-for-cloud-training SME) for the final step.

Anyway, I thought the expensive part was data acquisition and labeling? I like the 'surrogate network' approach of learning a very-expensive-to-compute simulation or model, that doesn't need data, but the output of a costly simulation.

>Scientific logic is proving things by losslessly compressing statements to their axioms. Commonsense logic uses lossy compression, which makes it less accurate in edge cases, but also less brittle, more efficient, further reaching and more stable in most real-world situations.

EDIT: A simpler example might be helpful. We could, for example, train a network to recognize and predict orbital trajectories. Feed it either raw images or processed position-and-magnitude readings, and it outputs predicted future observations. One could ask, "does it really understand orbital mechanics, or is it merely finding an efficient compression of the solution space?"

But this question can be reduced in such a way as to made empirical by presenting the network with a challenge that requires real understanding to solve. For example, show it observations of an interstellar visitor on a hyperbolic trajectory. ALL of its training data consisted of observations of objects in elliptical orbits exhibiting periodic motion. If it is simply matching observation to its training data, it will be unable to conceive that the interstellar visitor is not also on a periodic trajectory. But on the other hand if it really understood what it was seeing then it would understand (like Kepler and Newton did) that elliptical motion requires velocities bounded by an upper limit, and if that speed is exceeded then the object will follow a hyperbolic path away from the system, never to return. It might not conceive these notions analytically the way a human would, but an equivalent generalized model of planetary motion must be encoded in the network if it is to give accurate answers to questions posed so far outside of its training data.

How you translate this into AlphaFold I'm not so certain, as I lack the domain knowledge. But a practical ramification would be the application of AlphaFold to novel protein engineering. If AlphaFold lacks "real understanding", then its quality will deteriorate when it is presented with protein sequences further and further removed from its training data, which presumably consists only of naturally evolved biological proteins. Artificial design is not as constrained as Darwinian evolution, so de novo engineered proteins are more likely to diverge from AlphaFold's training data. But if AlphaFold has an actual, generalized understanding of the problem domain, then it should remain accurate for these use cases.

For example specific points on this object are a folded proteins. I suppose then it is how well does this get encoded, does it know about "properties" of this surface, or is it more like a rough kind of point cloud because you have sampled enough and then it does some crude interpolation. But maybe that does not respect the sort of properties in this object. Maybe there are conservation laws, symmetry properties, etc which are actually important, and then not respecting that you have just produced garbage.

So I think it is important to know what kind of problem you are dealing with. Imagine a long time scale n-body problem with lots of sensitivity. Maybe in a video game it doesn't matter if there is something non physical about what it produces, as long as it looks good enough.

Maybe this interpolation is practical for its purpose.

But I think we should still be careful and question what kind of problem it is applied to perhaps. Maybe it's more like a complexity vs complicated question.

> does it know about "properties" of this surface, or is it more like a rough kind of point cloud because you have sampled enough and then it does some crude interpolation

Say that there existed some high-level property such as "conservation of energy". A "knowledge system" which learns about that property would be able to answer any questions related to it after reducing to a "conservation of energy" problem. Is the same true for NNs? The way folks talk about them, they sound like they can compress dynamically, and would therefore be able to learn and apply new high-level properties.

Also, do NNs have "rounding errors"? We have confidently learned that energy is conserved, but would NNs which never had that rule directly encoded understand conservation as "exactly zero", or "zero with probability almost 1", or "almost zero"?

I'm similarly disappointed that, like with that effort, the methods and techniques will not be shared with the scientific community.

[0] https://deepmind.com/research/open-source/alphafold_casp13

[1] https://twitter.com/demishassabis/status/1405922961710854144

There are tons of science problems where there are just not enough gray matter because it's just too expensive to train scientists. ML can crunch any data and result and speed up research by guiding experiments, where normal research just doesn't have enough resources to do so.

Of course, it really only works if the scientists are able to understand data and how to use ML, which is why computing becomes just a tool for a scientist, nothing else.

And again, ML is not really "smart", it's just sophisticated, improved statistical methods.

>Of course, it really only works if the scientists are able to understand data and how to use ML, which is why computing becomes just a tool for a scientist, nothing else.

Ideally in science you would like to use literally anything else other than ML if possible, fitting models come with their own challenges and neural networks are even more of a nightmare. Understanding the world well enough to hard code a rule is always preferable to fitting to some data and hoping the model will come up with a rule. While there has been some attempts to use ML for feature detection it then takes a lot of experimenting to generally show if it detected signal or just some noise in your data.

Most of the things that would accelerate science would either require AI much more complex than we currently have (basically replacing lab assistants with AI) or are incredible research undertakings in their own right like Alpha Fold, Deep Potential Neural Networks, etc.

While AlphaFold 2 is a tremendous achievement, to me the major drawback is the blackbox approach. It means it is very difficult to know when the model is outputting garbage and it also doesn't directly lead to new insights.

A much more interesting approach: "Discovery of Physics From Data: Universal Laws and Discrepancies" [1]

If ML did that, then it would be much more interesting.

[1] - https://www.frontiersin.org/articles/10.3389/frai.2020.00025...

Ie., ML makes the assumption that data points are IID.

The whole purpose of science is to produce models which explain why data isnt IID.

You can also separate cause from effect by using causal inference, under some assumptions.

> ML makes the assumption that data points are IID.

Common ML algorithms do, but it is done for practical reasons rather than a limitation in the mathematics.

> The whole purpose of science is to produce models which explain why data isnt IID.

And ML can greatly help in this, though it is not a silver bullet.

Causality involves a deep understanding of a phenomenon in science.

For example, the standard model of physics is pretty good at describing the real world in a good enough manner because we understand a lot of it. The difference with correlation and causality, in my view, is human, scientific understanding of what things are. Formulas, data or drawing are not enough.

For example there might never be a way to prove natural selection, even if there is a lot of data available, but a lot of scientific consensus is enough to describe causality.

Science is derived from scio, which in latin means knowledge.

It's true that in a way, ML allows new things, but which are still obscuring real knowledge...

I'm still curious about analysis of trained networks.

You know that ML is a really important part of a lot of companies that aren't "Science", right?

Like Google search result rankings rely on ML.

It's clear to see how ML is useful for science, but why exactly do you think it's *only* useful for science? It seems like in order for that to be true, you'd have to expand the definition of science to basically everything.

https://paperswithcode.com/sota

> Neither the Oxford Protein Informatics Group nor I accept any responsibility for the content of this post.

And just stopped. I get it that you don't want to speak for your employer or academic institution but the "nor I" part is just weird.