AI Is Ushering in a New Scientific Revolution
thegradient.pub
thegradient.pub
That's the first in a series of bs claims. AI tools can, in time, bring an engineering revolution but a new scientific dark age whereby idiots with little to no domain knowledge use "ai" tools such as the super resolution ones mentioned, draw wrong conclusions and then publish them so other idiots can do flawed ai-powered (or not) meta-analysis and further damage the state of concensus and integrity in their field.
I have in fact been involved in projects that did exactly what you claim is BS - read a large corpus of scientific papers (and written medical records), and discerned connections worth further scientific investigation that had not been detected by the body of scientists following the field, ultimately leading to new drug discovery science.
So while you are right that mateo1 also did not prove his claim, your experience doesn't disprove his claim.
It is called an Artificial Intelligence that engineering product that can replace a Natural Intelligence in finding solutions - such as, through the automated development of sophisticated function approximators. This is one thing, and akin to the content of the article. And it is one meaning of the term "intelligence" ("intelligence₂").
On the other hand, I have seen algorithms that appear to collect associations by frequency being called Intelligence, in a very different sense of the term, and to that I note:
-- that Intelligence₁ is an ontology developer - in my terms; it "investigates what things are". In the very recent terms of James Fodor, which I present here because it is a different formulation by coincidence just published today, «humans learn by making structured mental concepts, in which many different properties and associations are linked together» [actually the crucial part is in /how/ the concepts are refined];
-- and that without operation of "critical reflection" crucial in the ontology development, and instead with a "tendency" to just trust the input and solve conflict without foundational considerations, an inputs-imitating system, which may be said "based on parroting", is the opposite of Intelligence₁ .
Strong disagree.
... but I'm having a hard time seeing any way to prove that I'm right. I think it may be that my working definition of "reflection" differs from yours.
And by the way, explicit critical thinking is expected to be, as it was in some regions, taught in primary school, e.g. behaviourally when children are taught that their peers being just six-years-old, they must be wary in their interaction, using their own judgement if only for safety; e.g. scholarily when they are not just told that '1+1=2', but made see that two items make a couple (and only then made learn menmonics like tables). The «field of expertise» of these "«trained professionals»" is just daily mind operation in general.
Of course, implicit critical thinking is just the operation of conceptual revision that would anyway go on unconsciously. And it does: people come up with "ideas".
As Prof. Patrick Winston insisted, "«adjustment by feedback»" is virtual, implicit or explicit mental experiments. You do not need to jump out of that window to know something about the consequences.
Any entity that repeats a prejudice - which is a clearly consequential and even studied effect of parroting - without "looking into it", internally discussing it, is unintelligent in the defined sense. If you have an idea that some phenomenon can be emergent from some mechanism, there will have to be foundational considerations to be proposed for it.
Parrots just repeat simple phrases, these language models recombine words to fit new contexts in ways that make sense. They memorise facts and common sense knowledge and use it to successfully solve hundreds of language based tasks. They are problem solvers, not parrots. This term was invented by an anti language model activist to discredit the big language model approach. Fortunately people still forge ahead with research and we get the likes of GPT-3, PaLM and DALL-E.
[1] Sutton's "Bitter Lesson" - the importance of scale in AI models. Even AI researchers don't like to accept this fact. No wonder people are incredulous.
http://www.incompleteideas.net/IncIdeas/BitterLesson.html
Richard Sutton is one of the top AI researchers with 115,498 paper citations. https://en.wikipedia.org/wiki/Richard_S._Sutton
The page from Sutton makes a statement that it is an error to implement imitations of human mind, especially hardcoding in models of it. The implementation of Intelligence is not an imitation of human mind, nor it needs hardcoding in simplified models.
> problem solvers, not parrots
They may solve language based tasks, yet Intelligence is a different matter. Parroting was used here in the sense of "repeating acritically inputs", not in the sense you used it.
> invented by an anti language model activist
And others, as you can see from this case, coincidentally decided for that term independently and unaware that others used it (as is normal). Probably driven to that having seen somebody call "intelligence" what is described as a system of consolidated associations - and noting that Intelligence is the opposite.
And Magic: the Gathering is Turing-Complete [1], but we are not communicating over an implementation of the http protocol in Magic: the Gathering.
Because that would be awfully inconvenient.
______________
Pretty much any way you turn you'll find classes of universal function approximators. The almost 200-year old Weierstrass-Stone theorem says that polynomials have this property. More prosaically, locally constant functions have this property (that's how the Riemann integral is defined). Trees and random forests exploit just this fact. Trigonometric polynomials have this property (via Fourier series). Wavelets.
Now, here's a mind-blowing fact: did you know that addition and 1-d functions form such a class of universal approximators. It's called Kolmogorov-Arnold theorem [1], and it's the origin of the generalized additive models (GAM) in statistics.
Back to neural networks: the universal approximation theorem is valid with a single layer. But nobody uses them with a single layer. Which means, people don't use NN's because of this (quite useless) property. Separately, this theorem is just an existence theorem, and it's not concerned with the complexity if the approximation. The approximation eventually relies in some form of optimization (see all the flavors of stochastic descent). But optimization, even convex optimization, is known to be NP-hard.
So, these are all good reason to discard the universal approximation theorem when talking about neural networks.
[1] https://en.wikipedia.org/wiki/Kolmogorov%E2%80%93Arnold_repr...
Yes, that's right, but the problem with trying to approximate a function you don't know is that you don't know how well you've approximated it. You're left having to, well, approximate, the accuracy of the approximation.
That's why everyone in machine learning trains their models and then tests them on a held-out test set. In theory, the error on the test set should predict the error on the true distribution (and that's under PAC-Learning assumptions). Unfortunately the way this is done in practice is sloppy and uninformative. For example, researchers usually have access to the test set during training, they split the training/test set to 80/20% so that the error calculated on the test set does not generalise well to truly unseen data, etc.
Bottom line: all this approximation stuff is great in theory, but not so good in practice. Or, rather, in theory machine learning systems have almost perfect accuracies, in practice we have no idea how good they are.
What can happen is that people train various models, then run them on the test set and decide that one or two of them are promising, and rinse and repeat. This creates an inefficient optimization (mediated by the researcher) on the test set. BTW this is absolutely unacceptable practice.
And it is as absolutely unacceptable practice as it is widespread, which is to say, almost universally so.
And the reason for that is the poor generalisation ability of neural nets, which basically forces the hand of researchers: splitting training/test sets by 20/80% would give a much better approximation to true error, but that would leave too litle data for actual training. So people just train with as much data as they can and hope, or don't care, about the true error.
This kind of thing is so widespread that not only is it standard practice, and accepted by the majority of the field, but, I suspect, a majority of people in the field don't even realise how little sense it makes. It's just how people learn to do this stuff and they never question it because the graphs look good and the papers get cited.
We're stuck in a very broken paradigm today in machine learning and that's all because we optimise for self-selected metrics of success. I find it extremely concerning to see the same paradigm applied to science. We risk destroying our ability to understand the world irreparably, or at least for such a long time that we will never recover from it.
> very broken paradigm
Well, that first mention makes it evident that the context is "Big Data driven Machine Learning", and this paradigm linked to BDDML is not the "intended ML" - it is just a passage in our toying. Because it is also very clear to us (well¹) that some proper learning has "one shot" properties, so at least until single information instances are determinant in model building (as opposed to slightly retouching weights) like in symbolic communication and more, we will know that something is amiss, that we are still crossing the river's fords. But the preparation of the environment to enable e.g. "one shot learning" is the current problem in such direction.
(¹even though some seem as if so inebriated with the current toys to claim "that's it, done, it rolls, TheBall 2.0, it's animated, it lives, we have created the Golem ... Look, a gramophone, it speaks, ask him for advice and serve him food")
This given, it is not that the "toying" has not produced noteworthy aids in real life application.
> I find it extremely concerning
That concern is that we are in need of enhanced human learning - we have more and more complexity to manage -, yet the effect and promise from these areas swing from boost to complacency. Some ask, "where is the problem in the "transparency problem", in blackbox responses": in that we do require more understanding, the human part is that which should primarily seek refinement.
https://github.com/stassa/louise#capabilities
In particular the second example listed there. A trivial example, but one that cannot be reproduced by current approaches without big-data pre-training.
I bring up the 80/20% training/test split that is standard in machine learning because I remember an interaction with my supervisor at the start of my PhD. In one of our meetings my supervisor asked me about the details of some experiments I was running, with a system called metagol (linked from the Louise repository above). En passant, I mentioned that I was training with a 20/80% training/test split and my supervisor stopped me to ask me if I thought that was a standard setup for machine learning. Thinking he meant the splitting of my training data to training and testing partitions I answered, a bit bemused, that yes, of course, that's the standard thing. To which my supervisor laughed and replied "I don't think so". Later of course I realised that he meant that the done thing in machine learning is to use most of the data for training and leave as little as possible for testing.
In Inductive Logic Programming, it's typically the other way around, and the training sets are often a few examples, like a dozen or so. Of course our systems don't do the spectacular, impressive things that deep learning systems do, but then again deep learning systems aren't anywhere near ILP approaches for sample complexity and robust generalisation from very few examples (as in one). And of course we don't have a dozen thousand graduates racing to out-do each other with new feats of engineering. Which is a bit of a shame because I think that if we had a couple more hundred people working on ILP, we 'd make huge progress in applications, as well as in understanding of machine learning in general. And the resulting systems would address many of the current concerns (mine anyway) about the use of machine learning in science (e.g. no more black boxes; "ultra-strong machine learning" in Donald Michie's sense of a system that improves the performance of its user; etc).
Oh well. It's probably all for the best. Who wants to build genuinely useful and intelligent systems anyway?
I do :) We need all the intelligence we can squeeze in this historical phase (where, among the rest, some conditions enabled an influent propagation of the opposite).
And: I will never be able to read the whole of even just my personal library. I dream of a "son" with the ability to do it... :)
The idea that the universal approximation theorem is something special about neural networks is completely wrong.
Deep learning is valuable because it represents a class of functions closed under composition. That allows it to discover features at different scales.
You can in principle add composability to other classes of functions. You would end up with deep splines, for example. Or deep radial basis functions, or deep wavelets. And people have tried those, but it just didn't take off. It was probably the advantage of the first mover. Deep NN's turned out to be useful, people built a large ecosystem for them, and then other deep things (like deep splines or wavelets) just did not show a significant advantage. Yet.
It was probably because of the "inspiring" stretched simile between nervous system and network of non-linear operators.
Also, there is a matter of the efficiency of learning algorithms: would the other classes of functions that satisfy the universal approximation theorem also use gradient descent and similar and subtricks, or would they allow different tricks for their adaptation to the desired results?
> It doesn't change what I said
Well, you stated that «people don't use NN's because of this (quite useless) property». I replied that the general idea of automatic function development may drive towards this promised Eldorado that would not only spare us from the burden of solution definition, but also (with the right tricks implemented) possibly confide such solutions to us in a "spoiler".
Of course, it can only answer with mechanisms that are identified in the literature it analyzed - it's not a PhD biochemist / M.D. able to infer or speculate on mechanism from first biochemical and medical principles. But there is no way to interpret what it can do, in my mind, other than to say that the model has learned drug biochemistry at some significant level, and can represent that learning to a user.
In the case of what Deep Mind has done with protein folding, I don't know how you interpret its capabilities other than to say their model has learned much of the fundamental body of rules and facts that govern how amino acid sequences role up into proteins. There is a very real sense in which it embodies a theory of protein folding, much as say, Maxwell's equations and a knowledge of calculus embodies a theory of electromagnetism. That it cannot as yet articulate the rules of that theory is clearly a profound limitation, but it's a limitation of conveyance, not of utility.
But your last sentence... if it had the rules programmed in, shouldn't it have also programmed in the ability to name the rules? If it learned them from reading papers, shouldn't it know the terms from the papers? Or is the problem that it doesn't have at all the ability to trace out the chain it used to reach the conclusion? Is the problem not so much expressing the chain as much as it is knowing the chain?
Deep Mind Alpha Fold did not get the rules from reading papers. It uses a combination of learned (from being trained on known protein structures) prediction of structure, and iterative application of learned refinement, constrained by physical and chemical rules. So, some of the knowledge required is "programmed in" as you suggest, but most has been learned by the algorithm through training on known structures. My point was that the result is nevertheless a workable effective procedure for predicting protein structures.
The other system that I referred to did learn from reading literature. It is a fundamentally different AI architecture (a variant of transformer architecture for large language models). It has no physical knowledge programmed in.
You know, for me it is clear that the ability to retrieve information is not the same as the ability to understand the retrieved information.
For example, you could store a corpus of scientific texts in a database and then retrieve pieces of it with a query, but that doesn't say the database understands the corpus.
And this is relevant to your comment because with transformers we still have essentially databases of information that can only be retrieved by the right query, where the information is stored as weights in the trained model and the query is a suitable prompt. In fact, prompts to transformers are so finicky and must be tailored so carefully to retrieve the desired piece of information that they are not very different than a SQL query, albeit one in natural language.
I'm offering this as a counterpoint to your assertion that "the model has learned drug biochemistry at some significant level". What I understand from your comment is that your model has learned to represent a body of text, nothing more, nothing less.
For example, we can see that clearly in chess and other board games. Chess engines play by executing some variant of the minimax algorithm and they can out-play humans. But human players don't typcially try to run minimax in their heads when they play and in any case can play chess without trying to emulate a chess engine. The same goes for image classifiers (it is extremely unlikely that we match sets of pixels to numerical labels in our heads, to the point of absurdity) and for language models (again it's absurdly unlikely that we calculate the probabilities of sequences of tokens).
I think what you say comes from an assumption that because humans exhibit certain behaviours, and humans have human intelligence, any system that exhibits similar behaviours must also have human intelligence. Which, if you look carefully, does not follow.
Anyway historically people have always been quick to identify the capabilities of computers as artificial intelligence, just because we tend to think of anything to do with computation as requiring intelligence. But we had machines that could perform calculations automatically, through entirely mechanical procedures, for more than 200 years now, and most of those we don't think are artificially intelligent. For example, I don't think anyone would consider the Pascaline to be an AI, and yet the Pascaline can do better arithmetic from any non-human animal, and from many human animals, also. So, does arithmetic require intelligence, or not?
How much of what we can do thanks to our intelligence can also be done by a dumb, mechanical automatism? This I think is the important question in AI research today and one that, if we could answer it (which we can't) would finally put the lid on wishful thinking about LLMs exhibiting "a different sampling" of intelligence than ours.
Would not you say instead that «model has learned /given statements about/ drug biochemistry at some significant level»? You say it yourself that it cannot «infer or speculate on mechanism from first biochemical and medical principles».
Students should not pass exams because they have memorized the book - they would be said to "have memorized the notions" and not to "have learned the matter". (We may note that, linguistically, 'learn' has a strong connotation of activity: it is "track chasing", so "investigating".)
> it embodies a theory of protein folding
This reminds me of the surprise researchers got from some AI (possibly even classical - simulated annealing?) produced models for microprocessors, where the "monstrous" (unordered) layout was discovered to have used important non-processing chunks of immediately unexplicable circuitry just for thermal properties and timing. The AI can come up to an effective solution, which can be theoretically very much not clean.
In particular, the protein folding example is extremely important and validates that AI is indeed having a big and real impact on science.
I sure as heck can't keep track of all the blogposts and articles that come out every day about software development. But I can expertly wield "an AI" to answer my questions and do just-in-time research when it's needed to solve problems.
For a more measured view on the use of AF in drug development (especially), check out med chemist Derek Lowe's blog, especially his articles: "Fooling the Protein Folding Software" https://www.science.org/content/blog-post/fooling-protein-fo... , and "AlphaFold Excitement" https://www.science.org/content/blog-post/alphafold-exciteme...
"SCIBERT: A Pretrained Language Model for Scientific Text" https://arxiv.org/pdf/1903.10676.pdf
"Domain-specific language model pretraining for biomedical natural language processing" https://dl.acm.org/doi/pdf/10.1145/3458754
These are just examples I pulled up in a couple minutes of Googling. Yeah, the wording here may be a little too grandiose, but nevertheless the fact that AI can read through vast sets of papers and data quickly and help humans sort through it is fundamentally true.
I'm asking because a scientific paper's purpose is to make one or more claims (but usually one central claim), that are supported in the paper by theoretical and empirical results. It is hard to know whether any particular claim is true, even when reading the paper carefully and scrutinising, or even attempting to replicate, its results - let alone otherwise.
That seems similar to the idea of self-driving cars that screw up frequently and have to suddenly turn over control to the driver being safer than just paying attention to driving.
Peer review sucks now. Reviewing something that makes errors that no human would make seems like it would be a lot worse.
This stresses the call for humans not to be lightminded and decadent: new instruments are around, and they require to be managed competently and properly.
It's been overhyped so much that it's very hard to convince non specialists that basically all the framing of AI they can find is hype. The best I can say is, show me a company that's actually making money off of a product that's based around one of these claims, that actually has the ML as a core differentiator.
No doubt people will still find headlines, but they'll be more scarce, and if you look into the commercial success of such companies, you'll find almost nothing.
Disclaimer, I run an AI company and am an AI bull. Misplaced hype helps charlatans and scammers
For instance, asteroid discovery. Yes AI can be used to discover asteroids. BUT every asteroid needs to be checked by a human and some will undoubtably be false positives. We also miss some unknown true positives.
We often (depending on method) can’t explain why [in-fact] why something is a true positive vs false positive from the network itself.
Which is why I say AI doesn’t make discoveries. It’s not identifying something knew, it’s identifying something that looks like what it’s seen before. It’s error rates in the scientific frontier is also going to be unknown; as we often lack data. Using AI could potentially limit the scope of new discoveries - depending how heavily it’s relied on by hyper optimizing on the wrong items.
Anyway, sorry for the rambling, as a research in deep learning I find this trend something to monitor.
As for whether finding patterns is discovery, I feel very whatever about it.
Here is the study:https://www.thelancet.com/journals/landig/article/PIIS2589-7... When i said insanely i didnt remember correctly, just saw even if reduce to 4x4 pixels they claim a statistical difference. So mind blowingly ludicrous resolution might have been a better term.
I'm pretty sure forensic anthropologists have been doing that for decades.
However, I suppose there's a possibility we could become over-reliant on it and blind to the sorts of new ideas that you'd generate from manually exploring the data set yourself.
For example, if the state-of-the-art AI asteroid discovering program has a blind spot where it never detects a certain type of asteroid, we might stop finding that type of asteroid altogether. Then when the training datasets themselves are mostly derived from asteroids that the AI had previously found, any blind spots and biases could end up self-reinforcing.
Then again, is this really all that different to the many human-driven blind spots science has had in the past?
No. Humans are fallible like AI. Most humans can't perfectly apply logical reasoning to common day situations.
For example, in reinforcement learning, rewards are provided by the environment, but there is no point where humans give labeled examples of what constitutes good play.
For an even more extreme example, in evolutionary learning there is no reward provided to the algorithm and there can be no training data visible to the algorithm. In effect, it manages to work as an optimization because sometimes there is a lack of data - something dies before reaching a point where it should produce another generation.
If we were only bumbling fools that stumble in the dark we'd still be stuck in dark caves without fire. But we can be -and we often are- much more than that.
But not our machine learning systems. Those are stuck in making stupid mistakes and not being able to generalise away from them. For the time being.
You can ask GPT-3 a question and it may come up with a smart answer or a dumb answer, and it won't know which is which. But we will.
Because, yes, we can make dumb mistakes but we can also understand that they are mistakes, and that they are dumb, and why they are mistakes and dumb, and we can recover from them and stop ourselves from making the same mistakes. We have blind spots, sure, but we can see through them, eventually, with hard work, after a very long time, very haphazardly.
But not our machine learning systems. For our machine learning systems the entire world is one big blind spot.
That is how humans are different to machine learning systems.
Yes, because humans can build models that are explanatory as well as predictive, and so can, eventually, get over blind spots. See for example the theory of epicycles and how it was eventually superseded by Kepler's laws and Newton's.
The danger with deep learning for science is that we will again be stuck in a rut, relying on (what we think are) good predictive models without explanatory power. In the case of astronomy and the theory of epicycles, the rut lasted more than a thousand years.
That is a terrifying thought, considering the mess we currently find ourselves in and the urgent need for true scientific understanding of the world, before we destroy it.
In Go, there are hundreds of years of tradition studying the games of past master players, and analyzing how their innovations changed our cultural understanding of the game. Humanity can't have it both ways. If those contributions were valuable when made by a human, they should be equally valuable when made by an AI.
[1] https://www.amazon.com/Using-AlphaGos-Enlarged-Corner-Enclos...
This took multiple human generations for us to accomplish, with literal geniuses applied to the task and spending their whole lives on it. Rediscovery and expansion happened in a wall clock time that is better measured in minutes than generations.
[1]: https://www.amazon.com/Game-Changer-AlphaZeros-Groundbreakin...
There’s conflating of terms here.
I have no interest in jumping down a semantic rabbit hole with you lettergram (are you just trolling?).
As a Go enthusiast with an entire bookshelf dedicated to books on Go like: Invincible, The Games of Shusaku, and 400 Years of Go in Japan, I'm comfortable stating that in my opinion the new moves popularized by AlphaGo mark a cultural shift in the game that is as significant as any of the big names in the history of the game.
Take for example the wonderful Wikipedia article on the innovations of Go Seigen[1]. Seigen's innovative opening plays in the 1930s is one of the reasons he is considered by many to be "the greatest Go player of the 20th century". [2]
Again, my point is this: Go Seigen's beautiful discoveries on the Go board made him a Hero to generations of Go players.
We can debate whether AlphaGo's discoveries are equal or lesser than Seigens, but they are within the same order of magnitude: both players triggered a once in a generation sea change in how Go Culture thinks about the game.
To brush aside AlphaGo's discoveries, would be to brush aside Seigen's as well. Which is demonstrably wrong.
So, fine. Is "discovery" the wrong word for what Alpha Go has achieved? Whatever. I don't care.
Lettergram, If you really have a point, and you really care about convincing me of your point, then reframe your argument in such a way that clearly differentiates Go Seigen's accomplishments in a meaningful way from those of AlphaGo. In what way does Go Seigen's legacy stand apart from AlphaGo's?
[1] https://en.wikipedia.org/wiki/Shinfuseki [2] https://en.wikipedia.org/wiki/Go_Seigen
Yes humans can be used to discover asteroids. BUT every asteroid needs to be checked by a human (or AI) and some will undoubtably be false positives.
We often can’t explain from the brain itself.
Humans don’t make discoveries.
(I guess you get my point.)
Here's the gist: A human can show that 50 people are dying of cancer because they were recently exposed to toxic sludge.
An AI can't say a bunch of people were exposed to toxic sludge - it can only raise your prediction of getting cancer because you live near such and such coordinates (the coordinates where toxic sludge exist - but that's not visible to the AI)
If those are reasons ai don't make discoveries, and they all apply to humans as well, then it must be that humans also don't make discoveries.
This is obviously false, which refutes the parent's point.
The asteroid has to be checked by a human (or AI) part is incorrect because a computer AI that's checking just images or data can't (yet) call a team of astronomers to focus telescopes somewhere and double check it. That part would almost nearly require general AI. Yes, someone could write a single-use software that tells a bunch of astronomers to check something that AI has found, but that's NOT "the AI" doing that.
Humans don't have "training data" they have something else. I don't think "training data" and that "something else" is the same (yet).
What do humans get instead of training data? What do you mean by "something else", and why can't some artificial neural network (ANN) not have the same? Maybe you mean it does not get supervised labeled data? But unsupervised data is also just training data, and in fact ANNs are very successful in also using such unsupervised training data.
So I really do not see the distinction between AI and human intelligence. And the argument why an AI can't make a discovery and a human can.
I haven't been reached out by OpenAI 5 in Gmail saying hey do you want to play a game of Dota 2 with me? It hasn't called me on the phone wanting to hang out. It's completely different.
When those things happen we can talk but the original poster and myself are predicating this on "CURRENT AI" not theoretical AI.
What you did, when you pointed out that his obviously wrong conclusion was wrong and asked him to defend it in order to be proved right is called a straw man. His point was that the conclusion was wrong and so insisting he defend it makes it very clear that you have poor comprehension of the logical tradition in which his comment is rooted.
Also, you're jumping to conclusions. The OP did something that is indeed very common on HN. You concluded that what they did was based on an ancient tradition of logic. Quite a big jump there.
Who cares if they aren't people from ancient Greece? I said it was a tradition rooted in logic, not that they were from Greece. As educated people in the modern world, we are taught about logic. You should know this better than most; a lot better, actually. Which goes to show how reasonable it is that people on this forum would be expected to understand points rooted in logic.
That said, it is Aristotle you ought to have referenced, "Aristotle sets the stage for all later theories of analogical reasoning. In his theoretical reflections on analogy and in his most judicious examples, we find a sober account that lays the foundation both for the commonsense guidelines noted above and for more sophisticated analyses." [1] "Aristotle proves invalidity by constructing counterexamples. This is very much in the spirit of modern logical theory: all that it takes to show that a certain form is invalid is a single instance of that form with true premises and a false conclusion. However, Aristotle states his results not by saying that certain premise-conclusion combinations are invalid but by saying that certain premise pairs do not “syllogize”: that is, that, given the pair in question, examples can be constructed in which premises of that form are true and a conclusion of any of the four possible forms is false." [2]
> Also, you're jumping to conclusions.
The alternative conclusion is that the person is too stupid to realize that obviously false claims are false, but simultaneously believes that making false claims with bad logic will mean that when they say "you get my point" they will be found to be convincing. That is an extremely implausible conclusion. It seems to me that the way to get to that conclusion is to dismiss people with prejudice on the basis of "I logic, they rhetoric" and to me at least that seems against the spirit of the Hacker News guidelines. So I don't think I made a big leap. I think I chose to assume good faith.
[1]: https://plato.stanford.edu/entries/reasoning-analogy/
[2]: https://plato.stanford.edu/entries/aristotle-logic/#DisCouTe...
I was just providing more evidence about my reasoning that the use of counterexample via substitution really does go as far back as ancient Greece, why that still applies to modern people and explaining why it seems more plausible to assume the original poster was employing a refutation technique than to assume otherwise.
More or less. Is Pluto still a planet?
Historically, many good scientific models had generalizable predictive ability (relativity being a good example) that was used to make predcitions that led to discoveries). I would say that if a scientist can use a model (that's all an ML model is) to anticipate likely regions of phase space to explore (say, in making engineered glass or steel), it is in fact "discovering" things.
Also, when you reach the point of needing a human to double check all your predictions and the humans experts disagree about the golden label as much as the model has uncertainty in the prediction, the model has reached human performance, and "human checking the results" doesn't even make sense since they would increase the false positive or negative rate of the system. Already we are seeing some image recognition systems whose classificaftion of hard examples exceed that of the best human experts (while being hundreds of times fasteR).
By your logic, humans can't discover things, but AI can. You think you made the opposite claim - you didn't.
1. Consider that you can freeze a random seed, making an algorithm deterministic, but you can't do the same for a human.
2. Secondly, I dare you to explain to me exactly, down to the very activation of cells and synapses in your mind, how you came to understand these words as having meaning. You can't do it, but an AI that did this would do this implicitly. So it will do a much better job of explaining how this happens than you will.
So you haven't ruled out AI being able to discover things. You've claimed that humans can't, but AI can. It directly contradicts your thesis.
3. We can point to specific examples of novel discoveries produced by modern ML that were made without the use of good training data provided by humans - indeed, without any training data provided by humans. Moreover, these techniques weren't limited by their training data being bad - they were built around the idea that the training data they would generate was bad.
I think you're working with a very different definition of "explainable" than is commonly used in the machine learning literature.
It really shouldn't surprise you that we have to go back to the early parts of the field; his argument restated is "machines can't make discoveries, only humans can" and this confused thinking is something that gets addressed in the early writings much moreso than the latter writings.
Frankly, we've known algorithms as far back as Aristotle that were capable of discovery. Peter Norvig in AIMA cites Aristotle when talking about some search problems and path discovery is still discovery. Some of these algorithms literally include frontier in the algorithmic description; clearly a framing that implies thinking of them in terms of discovery. Honestly, what is really happening in my estimation is some people do this thing where they pretend AI is categorically reduced to the subset of things which they can criticize. It is a really weird tendency, but extremely common. You can see it in the persons post actually. Look at the contortions they go through to only focus on supervised learning - to only focus on a narrow subset of supervised learning restricted to deep learning - to only focus on a narrow problem domain of supervised learning - classification - and to only focus on a specific example of that problem domain (we can give good explanations for smaller models) - the ones that humans can't easily explain.
Uh, you do realize that AI and classifiers are two different things right? Like, there are literally, right now - like, already - AI which make a hypothesis and then run an experiment. Just because it is popular to phrase things as supervised learning problems and just because we often want to have classifiers as components of systems we build doesn't mean that is the only technique in AI that exists.
Honestly, your entire point reeks of this really unjustified hypocrisy that makes you think you don't realize this. You seem to pick the cherries for the human, but grab only the poisonous mushrooms for the AI. For example, you seem to close your eyes when you hear about the replication crisis in the social science acting as if humans only emit proofs when they don't, but you're more than willing to focus on how some subsets of AI don't emit proofs despite there being subsets of humans which don't emit proofs.
Let me give a somewhat broad example. In cases where a state space to be explored is absolutely massive and we have no currently good way to 1. prune it 2. traverse it to land at the states we want(that have certain properties), and so on. Then the best we have right now is "sample the state space." But "sampling" requires very good random number generation so that we're 1. not biasing ourselves (this is clear to everyone) but deeper than that 2. Good random number generation reduces the problem of us only grabbing states from the state space that confirm our existing ideas rather than the ones that might challenge our existing ideas.
How many scientists are developing their ideas on single workstations that are running out of entropy?
Lenses can be thought of thermodynamically as channeling the information included in rays of light into a different form more appreciable by humans.
AI is that for rays of data :)
* AI model’s insight helps astronomers propose new theory for observing far-off worlds https://techcrunch.com/2022/06/03/ai-models-insight-helps-as...
* Scientists Discover Method to Break Down Plastic in Days, Not Centuries https://www.vice.com/en/article/akvm5b/scientists-discover-m...
* Machine Learning Helps See into a Volcano’s Depths https://eos.org/editor-highlights/machine-learning-helps-see...
* Stanford University use AI computing to cut DNA sequencing down to five hours https://www.zdnet.com/article/stanford-uni-nvidia-use-ai-com...
AI is just a tool, and is increasingly becoming a big part of scientists' toolsets. I don't know if it's fair to say it will lead to a scientific revolution, but it's impacts will most certainly be vast.
Self-plug warning: I know of all of these articles because I run the newsletter Last Week in AI (lastweekin.ai) and co-host its related podcast (https://www.lastweekinai.com). We've covered a dozen or more of these types of stories, and new ones keep showing up. So it's definitely a big trend IMHO.
The first uses physics-informed neural networks in systems of ODEs to solve complex problems faster and even helps derive DEs that fit some data.
The second uses a fixed set of symbols to fit a tree of them to data. This gives a closed-form solution that best fits the data. Both are pretty hot in the Julia community and have results published
AI means we can literally brute-force any scientific mystery by trying out millions of algorithms until we find one that fits (like Wolfram's New Kind of Science). General relativity was a brilliant theory, but we don't have billions of Einsteins to keep coming up with new theories. Soon, we will. And obviously, brute forcing is only a last resort, AIs will be smarter than that for most problem-sets.
Who cares that it's a black box? Humans still need to verify a new model or theory (the AI will help with that too) but it'll shave decades off of discoveries.
Why would humans lose domain knowledge? (There's plenty of non-economists writing economics papers already, and they seem human!) AI might dramatically increase the rate of scientific discoveries, and make a lot more people want to become scientists. It'll also help reclaim science from big institutions and academia, and help small independent groups make significant discoveries again without gatekeepers. Patent clerks might be able to make scientific discoveries again!
And why would AI necessarily be purely imitative (i.e. only be capable of building off past knowledge, rather than generating new knowledge or at least using abstract reasoning)? See this tweet from a DeepMind researcher, plenty of people are working on this: https://twitter.com/Inoryy/status/1522621712382234624
To solve a specific problem, you just need a "mysterious solver". To increase human knowledge, you instead need a formally expressed insight.
In the article I read, «In 2020, Google’s AI team DeepMind announced that its algorithm, AlphaFold, had solved the protein-folding problem». Now, I understand that AlphaFold "solves protein-folding problem_s_", while "the protein-folding problem" would be a theoretical problem which calls for a theory to be solved. This is the difference. They are two different matters, having things done and having us more knowledgeable.
This is exactly wrong. If true, then AI could readily solve any human-level problem we could throw at it. But AI still can't master any of the basic skills inherent in human children, much less in adults working in construction crews, or health care aides, or janitors, or even just driving a car.
AI's ability to build robots capable of even basic tasks like giving a haircut or climbing over broken objects littering a stair -- they're nonexistent, and even after a decade of deep learning, they're not getting appreciably better.
Deep learning has shown us two things: 1) deep NNs can learn how to solve approximation problems astonishingly efficiently and capably. That's AlphaGo and AlphaFold. And 2) a lot of what humans assumed to be signs of intelligence, like playing go and chess, understanding voice and speaking, recognizing objects in images -- these are just "monkey see, monkey do" skills based in probabilistic pattern matching, which deep NNs can do very, very well.
But cognitive skills that require reasoning -- deduction, induction, and the recognition and rejection of contradiction -- these must identify relevant domain knowledge (semantics) and the ability to reason about it, not just the efficient matching of a pattern in a matrix using a GPU. Alas, the effective and scaleable means of solving these problems still remain a total mystery.
So if you can generate a zillion clueless hypotheses per second using the very best deep NN, it will still drive over a cop dressed like a clown who's waving his arms the path of your smartcar, nor will it ever invent general relativity.
The biggest effect it could have is ending science, convincing the powers that be that understanding is no longer a prerequisite of capability, and thus science can once again be relegated as a nerdy hobby that is of no interest to the state or corporation.
Anti-ignorance != anti-automation
Knowledge (of something) is something one could achieve, not something everyone must achieve. The relevant purpose is that those who should become competent and knowledgeable do so - for judgement and for production of further knowledge.
The proper analogy would be scientists have some big machine which tries different chemical experiments until it has a Covid test that matches the training data. But how would we even trust the training data in that case?
Having a good understanding of the basics of genetics and microbiology allows us both to understand how the tests work, and tests in the QAing of the testing procedure. There is no substitute.
Even if AI is really capable of doing so (I don't think it is), it is similar to saying that a different species (not human) will now do science for us. How useful is that for humans? If that were even true, it would mean that humans will be left in ignorance and start to rely (without ways to confirm it) on the results provided by this new species. In other ways, the whole hoopla about AI doing science is nonsense on its core.
Oh nice, thank you for that. Letting AIs doing science is domesticating ourselves. Shall the general AI feeds us kibbles too?
Is the title hyperbolic wrt this being a revolution? Maybe. Is it fair to say AI is 'ushering' a lot of new science research into existence? Absolutely.
"AI" is ushering in a new era of stuff that may or may not happen in the future.
We finally have great uses for all these transistors:)
It is like asking what makes something a cat, then coming up with a classificatory answer: oh, this image is a cat, that image, another cat, etc.