Comment by Isinlor | Hacker News Reader

Comment by Isinlor | Hacker News Reader

Parent Full thread

I'm not a biochemist. I don't know how much knowledge of how protein fold is important as opposed to knowing the final structure. My understanding is that current expensive method of X-ray crystallography in the range of $100k per protein also does not explain how protein fold, however the structure itself appear to be useful.

Many complex problems do not appear to be explainable in the way planet motions are explainable. The fact that many aspects of the universe are explainable in simple laws is rather extraordinary in itself. In reality, there may be just a lot of irreducible complexity. Maybe success of Deep Learning teaches us that certain things do not fit into framework of science that you assume to be the only correct framework. Reality does not care about what we would want to be true.

Maybe invention of AlphaFold will lead to better understanding of protein folding the same way that invention of steam engine lead to the laws of thermodynamics. Historically, science often lags behind progress of technology. Time will judge.

I also thought that the lack of resources will be your answer. In 2004-2009 neural networks were in quite similar position to ILP. However, NN showed significant potential as evaluated by the broad community at the time and so resources allocated to NN started to grow exponentially. ILP and symbolic methods more broadly do not show similarly convincing potential to the broad community. I'm sure OpenAI, DeepMind, Google, Micosoft etc. would not hesitate to dump equivalent resources they do into many other projects if they saw potential. DeepMind, Microsoft and Google especially hire people with very broad range of expertise.

What may be true is that the world is starting to reach limits of the resources that can be dedicated to deep learning. Exponential growth in resource use can not go forever unless deep learning will start feeding back into available resources and we are certainly not there, not yet.

> Yet the algorithms that I study run on a student laptop, are trained in a weakly supervised manner, take seconds to train, and generalise robustly from single examples without any sort of pre-training whatsoever

Show me that applied to: https://github.com/fchollet/ARC

>> Maybe invention of AlphaFold will lead to better understanding of protein folding the same way that invention of steam engine lead to the laws of thermodynamics. Historically, science often lags behind progress of technology. Time will judge.

It's not impossible but I'm concerned that neural networks will only be used as an excuse to not have to understand anything anymore (not just in biology, in all of the sciences). That would be a terrible outcome- a dumbing down of science and an eventual loss of our ability to understand the world. As you say, time will be the judge of that.

>> Show me that applied to: https://github.com/fchollet/ARC

I had a look at ARC back when François Chollet's paper came out that presented it. I was interested because a) it's the kind of problem that ILP eats for breakfast and b) Chollet's name is well-known and it would be a chance to get our work noticed by people who normally wouldn't notice it.

However.

ARC is proposed as a benchmark that is hard for current big-data approaches and that would need elements of intelligent reasoning to solve (if I understand correctly, that's why you brought it up yourself?). Such benchmarks have been proposed before, in particular the Bongard problems [1], in machine vision, and the Winograd schemas [2], in NLP. As with ARC, the "defense" of such datasets against dumb, no-reasoning, big-data approaches is that there are few examples of each problem task. And yet, in both Bongard problems and Winograd schemas, neural nets have now achieved high accuracy. How did they do it? They did it by cheating: instead of training on the original, and few, Bongard problems, for example, people created generators for similar problems that could generate many thousands of problems [3]. That way, the neural nets folks had their big data and they could train their big networks. Same for Winograd schemas.

... And the same thing has already happened, at a preliminary stage, with ARC:

https://arxiv.org/abs/2011.09860

In the paper I link above, the authors use a data augmentation technique that consists of rotations and colour transformations of problem tasks. It should be obvious that this adds no useful information that can help any attempt at solving ARC problems with reasoning, and that it only serves to help a neural net better overfit to the training tasks. And yet, the system in the paper achieves good results in a small selection of ARC tasks. Admittedly, that is a very small selection (only 10x10 grids and not on any of the held-out test set that only François Chollet has access to) but the important point is that it is possible to make progress in solving ARC without actually showing any reasoning ability, despite the claims of the Kolev et al. paper above; and despite Chollet's intent for ARC to avoid exactly that.

This is pretty standard in deep learning and the reason why I gripe about benchmarks in my earlier comment. Deep learning systems consistently beat benchmarks designed to test abilities that deep learning systems don't have. They do this by cheating, enabled by weaknesses in said benchmarks. They cheat by overfitting to surface statistical regularities of the data (see earlier cited papers) thus learning "shortcuts" around the intended difficulties of the benchmarks. The trained systems are useless against any other dataset and their performance degrades precipitously once exposed to real-world data, but that doesn't matter, because a team that trains such a system to a new SOTA result will get to publish a paper and claim its academic brownie points, and it's very difficult to argue with tis "results" without beating the benchmarks yourself, because of the publishing climate described in Hinton's quote, above.

Of course, if you want to beat the benchmarks yourself, you have two options. One way is to also cheat and throw a bunch of data at the problem without attempting to understand it or really solve it. The other way is to do it the "right" way, to try to design a system that really demonstrates the abilities the benchmark is supposed to be measuring. Which can sometimes be done, but it takes a lot of time and effort. In the case of ARC, that means learning core priors.

If you've read the Chollet paper that introduced ARC, there's a section on "core priors", that the author says a system must possess before it can solve the ARC tasks. Those core priors must be learned (or worse, coded in by hand). So anyone who decides to solve ARC the way its creator intended will need to spend a great deal of time teaching their system core priors, even before they can start on the actual problems. Meanwhile, some big, 30-person team at a large tech corp with a few milllion dollars' budget will be throwing a gigaflop of compute and a few terrabytes of data on the problem and "solving" it without having to learn any "core priors", by finding some way around the dearth of training data, just as was done with Bongard problems and Winograd schemas. Then anyone who has published any earlier results obtained "the hard way" will be left looking like a loser. Why would I want to subject myself to such a humiliation and drag the reputation of my field through the mud?

So, no, I'm not "showing" you anything. Sorry. If you're interested in the results of ILP and MIL systems, or my own work, check out the ILP and MIL literature. Ask me if you want and I can give you some pointers.

__________________

[1] https://en.wikipedia.org/wiki/Bongard_problem

[2] https://en.wikipedia.org/wiki/Winograd_schema_challenge

[3] https://link.springer.com/chapter/10.1007%2F978-3-319-44781-...

See Section 2.1 "The Dataset": "For each class of each problem we generate 20000 training images. We also generate an additional 10000 images per class per problem as a testing set."

P.S. This paper compares GPT-2,3 to Louise (the system I created for my PhD) and also to humans and Magic Haskeller (an Inductive Functional Programming system, learning programs in Haskell) on some programming tasks in P3 (a Turing complete language with only 7 instructions, kind of a precursor of Brainfuck). I'm not affiliated with the authors:

https://proceedings.neurips.cc/paper/2021/hash/0cd6a652ed1f7...

Louise does alright, outperforming all three other systems and humans on the "very high complexity" tasks, but GPT-3 outperforms it on simpler tasks. In truth, Louise can do better than that, especially with a newer version, not available to the authors, and that has improved one-shot learning capabilities. On the other hand, the comparison is unequal: I was able to find the programming tasks and their solutions in a github repository predating the publication of GPT-3, so GPT-3 most likely ingested and memorised them, while it probably also benefited by the many examples of similar programming tasks in Brainfuck, that can be found online. Louise, as usual, learned in true one-shot fashion, without having seen examples of the tasks or their solutions before.

I mentioned ARC because based on your description it should be "the kind of problem that ILP eats for breakfast". But based on what you say about required effort it does not appear to eat it for breakfast and requires significant engineering fine-tuning, so I call foul on your previous description of your work.

I agree, to a degree, with your assessment of the problem of cheating. Currently community is applying two fixes to that: 1) require few-shot learning ability 2) require good performance on many different few-shot learning benchmarks

My hobby is future forecasting and I was thinking for some time about creating a benchmark for AGI based on Bongard problems, ARC and Montezuma's Revenge. I did have a question about ARC that managed to surprise community: https://www.metaculus.com/questions/3762/will-early-2020-ai-... I have one about Montezuma's Revenge where I require some reasonable score after 30 min of game play, no specific to Montezuma's Revenge engineering allowed, not even retrieving rewards from internal game memory: https://www.metaculus.com/questions/5460/ai-rapidly-learning... I'm aware of the paper you mentioned around ARC and my fix to the benchmark would be to require few-shot generalization on task-level e.g. tested system, if it is a learning system, must have access only to couple of tasks for training before start of evaluation. The same with regard to Bongard problems. Requiring that all 3 benchmarks are solvable by one unified system IMO would be a reasonable test for general, visual intelligence.

DeepMind recently decided to include comparison of Gopher performance on Massive Multitask Language Understanding benchmarks vs. forecasters prediction. Forecasting communities seems to have gained some clout in AI circles (OpenAI is also cooperating with Metaculus on AI forecasting). Few-shot ARC forecast from Metaculus could give you an external estimation of the difficulty that you could use against reviewer 2.

Also, despite what you think about large language models memorizing everything few-shot SuperGlue is still standing: https://www.metaculus.com/questions/4932/when-will-ai-achiev... My intuitive evaluation of GPT-3 light text reasoning skills based on playing around with it matches few-shot SuperGlue score of 72 vs. 90 for avg. human.

> Meanwhile, some big, 30-person team at a large tech corp with a few milllion dollars' budget will be throwing a gigaflop of compute and a few terrabytes of data on the problem and "solving" it without having to learn any "core priors", by finding some way around the dearth of training data, just as was done with Bongard problems and Winograd schemas.

Compute is cheap and it reliably gets cheaper. Models like GPT-3 even with cost of tens of millions of dollars are economically justifiable. It's quite possible that OpenAI is currently turning profit on GPT-3. Real market performance is the ultimate benchmark of whether "performance degrades precipitously once exposed to real-world data".

>> I mentioned ARC because based on your description it should be "the kind of problem that ILP eats for breakfast". But based on what you say about required effort it does not appear to eat it for breakfast and requires significant engineering fine-tuning, so I call foul on your previous description of your work.

Yep, you caught me. ILP is not magick. You actually have to do some work to train ILP algorithms.

To clarify, there is no "engineering fine-tuning" required. ILP approaches rely on background knowledge, which can be thought of as a library of sub-routines from which target programs are composed. ILP is unique in that the representation of background knowledge is the same as that of target programs (they're all logic programs) so they can learn their own background knowledge incrementally. So the effort to solve ARC would include training an ILP system to learn core priors from some initial set of low-level primitives. Once learned, core priors would be included in the background knowledge needed to actually solve ARC tasks (along with other stuff like rotations, translations etc that don't have to be learned because they're straightforward to hand-code). Clearly, that would take time and work. Deep learning approaches simply replace time and effort with data and compute. But there is always a cost to be paid. You can't have learning for free.

Learning core priors is not going to be simple. First of all, we don't even know exactly what they look like. Chollet gives a list, which I think is not exhaustive, but there's no dataset with examples that can be used in training. So there's a lot of work to be done, even just to figure out how to learn the core priors. Or you could do like on Kaggle and solve a sub-set of tasks with some geometric manipulations and a brute-force search. But that, too, would be cheating and it wouldn't generalise well anyway.

>> Requiring that all 3 benchmarks are solvable by one unified system IMO would be a reasonable test for general, visual intelligence.

I don't think that addresses the real issue with benchmarks. Machine learning benchmarks don't work because we have no idea what we want them to test, and that in turn because we don't have a theory of intelligence and artificial intelligence. Benchmarks like the Bongard problems, or ARC, now, are just educated guesses about the kind of behaviour an intelligent system should exhibit. Because they are just guesses, we're not really going anywhere and every few years someone comes up with a new benchmark, claims that this one is better than earlier ones, and then watches in horror as some dumb machine rips it to shreds. This will continue for ever until someone comes up with a plausible theory of what, exactly, it is that we are trying to test. Chollet actually tried that, but I think wrapping it all up in yet another new benchmark removes the focus from the theoretical work and focuses it on racing for SOTA again, thus exposing his work to the same old big-data shennanigans.

For example, 72% accuracy on a language understanding task should indicate pretty strong language understanding ability already _if_ the benchmark was actually testing language understanding. 72% accuracy would certainly beat a dog, or a chimp, it would probably beat a toddler and may even beat an educated adult who is not a native speaker of the test language. So it's clear that testing "few" or "any" shot makes no difference because no language understanding ability is being tested anyway. The benchmarks are fundamentally broken. They won't get any better by tweaking them. You can't make a thermometer measure luminosity no matter how much you tweak it.

Anyway pre-training, as I noted earlier, removes all legitimacy from any claims of "few shot" learning. First we need to undersand what is being learned during pre-training and if that has anything to do with some element of intelligent, then we can decide what "few shot" performance tells us.

>> Real market performance is the ultimate benchmark of whether "performance degrades precipitously once exposed to real-world data".

That sounds like you're saying that a neural network model's performance has something to do with how well it does as a commercial product. You can't mean that, so please explain. What's the market got to do with anything?

P.S. Interesting stuff on Metaculus.

Reply on news.ycombinator.com