The Bitter Lesson (2019)(incompleteideas.net) |
The Bitter Lesson (2019)(incompleteideas.net) |
I think these lessons are less appropriate as our hardware and our understanding of neural networks improve. An agent which is able to [self] learn complex probabilistic relationships between inputs and outputs (i.e. heuristics) requires a minimum complexity/performance, both in hardware and neural network design, before any sort of useful[self] learning is possible. We've only recently crossed that threshold (5-10 years ago)
>The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin
Admittedly, I'm not quite sure of the author's point. They seem to indicate that there is a trade-off between spending time optimizing the architecture and baking in human knowledge.
If that's the case, I would argue that there is an impending perspective shift in the field of ML, wherein "human knowledge" is not something to hardcode explicitly, but instead is implicitly delivered through a combination of appropriate data curation and design of neural networks which are primed to learn certain relationships.
That's the future and we're just collectively starting down that path - it will take some time for the relevant human knowledge to accumulate.
"Massive, deep search" that started from a book of opening moves and the combined expert knowledge of several chess Grandmasters. And that was an instance of the minimax algorithm with alpha-beta cutoff, i.e. a search algorithm specifically designed for two-player, deterministic games like chess. And with a hand-crafted evaluation function, whose parameters were filled-in by self-play. But still, an evaluation function; because the minimax algorithm requires one and blind search alone did not, could not, come up with minimax, or with the concept of an evaluation function in a million years. Essentially, human expertise about what matters in the game was baked-in to Deep Blue's design from the very beginning and permeated every aspect of its design.
Of course, ultimately, search was what allowed Deep Blue to beat Kasparov (3½–2½; Kasparov won two games and drew another). That, in the sense that the alpha-beta minimax algorithm itself is a search algorithm and it goes without saying that a longer, deeper, better search will inevitably eventually outperform whatever a human player is doing, which clearly is not search.
But, rather than an irrelevant "bitter" lesson about how big machines can perfom more computations than a human, a really useful lesson -and one that we haven't yet learned, as a field- is why humans can do so well without search. It is clear to anyone who has played any board game that humans can't search ahead more than a scant few ply, even for the simplest games. And yet, it took 30 years (counting from the Dartmouth workshop) for a computer chess player to beat an expert human player. And almost 60 to beat one in Go.
No, no. The biggest question in the field is not one that is answered by "a deeper search". The biggest question is "how can we do that without a search"?
Also see Rodney Brook's "better lesson" [2] addressing the other successes of big search discussed in the article.
_____________
[1] https://en.wikipedia.org/wiki/Deep_Blue_(chess_computer)#Des...
I think the answer is heuristics based on priors(e.g. board state), which we've demonstrated (with alphago and derivatives, especially alphago zero) that neural networks are readily able to learn.
This is why I get the impression that modern neural networks are quickly approaching humanlike reasoning - once you figure out how to
(1) encode (or train) heuristics and
(2) encode relationships between concepts in a manner which preserves a sort of topology (think for example of a graph where nodes represent generic ideas)
You're well on your way to artificial general reasoning - the only remaining question becomes one of hardware (compute, memory, and/or efficiency of architecture).
Yes- because human players can only search a tiny portion of a game tree and a minimax search of the same extent is not even sufficient to beat a dedicated human in tic-tac-to, leta lone chess. That is, unless one wishes to countenance the possibility of an "unconscious search" which of course might as well be "the grace of God" or any such hand-wavy non-explanation.
>> It's possible that a search subnetwork gets "compiled without debugger symbols" and the owner of the brain is simply unaware that it's happening.
Sorry, I don't understand what you mean.
Some, but there's a LOT more context pruning the search space.
Watch some of the chess grandmasters play and miss obvious winning moves. Why? "Well, I didn't bother looking at that because <insert famous grandmaster> doesn't just hang a rook randomly."
Expert players have likely a very well-tuned evaluation function of how strong a board "feels". Some of it is explainable easily: center domination, diagonal bishop, connected pawn structure, rook supporting pawn from behind, others are more elaborate, come with experience and harder to verbalize.
When expert players play against computers, the limitation of their evaluation function becomes visible. Some board may feel strong, but you are missing some corner case that the minmax search observes and exploits.
Whatever human minds do, computing is only a very general metaphor for it and it's very risky to assume we understand anything about our mind just because we understand our computers.
My guess is that we're doing pattern recognision, where we recognize taht a current game state is similar to a situation that we've been in before (in some previous game), and recall the strategy we took and the outcomes it had lead to. With large enough body of experience, you can to remember lots of past attempted strategies for every kind of game state (of course, within some similarity distance).
I can certainly see how it could be considered disappointing that pure intellect and creativity doesn't always win out, but I, personally, don't think it's bitter.
I also have a pet theory that the first AGI will actually be 10,000 very simple algorithms/sensors/APIs duct-taped together running on ridiculously powerful equipment rather than any sort of elegant Theory of Everything, and this wild conjecture may make me less likely to think this a bitter lesson...
The first of anything is usually made with the help of experts, but they're quickly overtaken by general methods that lever additional computation
Yes, many AI fields have become better from improved computational power. But this additional computational power has unlocked architectural choices which were previously impossible to execute in a timely manner.
So the conclusion may equally well be that a good network architecture results in a good result. And if you cannot use the right architecture due to RAM or CPU constraints, then you will get bad results.
And while taking an old AI algorithm and re-training it with 2x the original parameters and 2x the data does work and does improve results, I would argue that that's kind of low-level copycat "research" and not advancing the field. Yes, there's a lot of people doing it, but no, it's not significantly advancing the field. It's tiny incremental baby steps.
In the area of optical flow, this year's new top contenders introduce many completely novel approaches, such as new normalization methods, new data representations, new nonlinearities and a full bag of "never used before" augmentation methods. All of these are handcrafted elements that someone built by observing what "bug" needs fixing. And that easily halved the loss rate, compared to last year's architectures, while using LESS CPU and RAM. So to me, that is clear proof of a superior network architecture, not of additional computing power.
Raw computation is only half the story. The other half is: what the hell do we do with all these extra transistors? [1]
In fact, is the researcher supposed to be building the most performant solution? This article seems alarmingly misinformed. To understand 'artificial intelligence' isn't a race to VC money.
Many researchers predict a plateau for AI because it is missing the domain specific knowledge but this article and the benefits of more compute that OpenAI is demonstrating beg to differ.
I do agree with the part about not embedding human knowledge into our computer models, any knowledge worth learning about any domain the computer should be able learn on its own to make true progress in AI.
> This, then, is the trillion-dollar question: Will the approach undergirding AI today—an approach that borrows little from the mind, that’s grounded instead in big data and big engineering—get us to where we want to go? How do you make a search engine that understands if you don’t know how you understand? Perhaps, as Russell and Norvig politely acknowledge in the last chapter of their textbook, in taking its practical turn, AI has become too much like the man who tries to get to the moon by climbing a tree: “One can report steady progress, all the way to the top of the tree.”
My take is that there is something intelectually unsatisfying about solving a problem by simply throwing more computational power at it, instead of trying to understand it better.
Imagine in a parallel universe where computational power is extremely cheap. In this universe, people solve integrals exclusively by numerical integrations so there is no incentive to develop any of the Analysis theory we currently have. I would expect that to be a net negative in the long run as theories like Gen Relativity would be almost impossible to develop without the current mathematical apparatus.
To play devil's advocate, I think retort to your comment about "intellectually satisfying" methods is "yeah, but, they work". And in any case, "intellectually satisfying" doesn't have a formal definition in computer science or AI so it can't very well be a goal, as such.
My own concern is exactly what Russel & Norvig seem to say in Hofstadter's comment: by spending all our resources on clmbing the tallest trees to get to the moon, we're falling behind from our goal, of ever getting to the moon. That's even more so if the goal is to use AI to understand our own mind, rather than to beat a bunch of benchmarks.
https://www.theatlantic.com/magazine/archive/2013/11/the-man...
At least a few of the original synthetic biologists are a bit disappointed in the rise of high-throughput testing for everything, instead of "robust engineering". Perhaps what allows us to understand life isn't just more science, but more "biotech computation".
It is exactly the point and it is something not a lot of researchers really grok. As a researcher you are so smart, why can't you discover whatever you are seeking? I think in this decade, we see a couple more scientific discoveries by brute force which will hopefully will make the scientific type a bit more humble an honest.
This seems problematic as a concept in itself.
Sure human players have a "human understanding of the special structure of chess". But what makes them play could be an equally "deep search" and fuzzy computations done in the brain that and not some conscious step by step reasoning. Or rather, their "conscious step by step reasoning" to my opinion probably relies on tops on subconscious deep search in the brain for pruning the possible moves, etc.
I don't think anybody plays chess at any great level merely by making conscious step by step decisions.
Similar to how when we want to catch a ball thrown at us, we do some thinking like "they threw it to our right, so we better move right" but we also have tons of subconscious calculations of the trajectory (nobody sits and explicitly calculates the parabolic formula when they're thrown a baseball).
I wonder what would be required to build a model that explores the search space of compilable programs in say python that sorts in correct order. Applying this idea of using ML techniques to finding better "thinking" blocks for silicon seems promising.
Oh, not that much. You could do that easily with a small computer and an infinite amount of time.
I recently participated in the following Kaggle competition:
https://www.kaggle.com/allen-institute-for-ai/CORD-19-resear...
Now, you can see the kinds of questions the contest expects the ML to answer, just to take an example:
"Effectiveness of movement control strategies to prevent secondary transmission in health care and community settings"
All I can say is that the contest results, on the whole, were completely underwhelming. You can check out the Contributions page to verify this for yourself. If the consequences of the failure weren't so potentially catastrophic, some might even call it a little comical. I mean, its not as if a pandemic comes around every few months, so we can all just wait for the computational power to catch up to solve these problems like the author suggests.
Also, I couldn't help but feel that nearly all participants were more interested in applying the latest and greatest ML advancement (Bert QA!), often with no regard to the problem which was being solved.
I wish I could tell you I have some special insight into a better way to solve it, given that there is a friggin pandemic going on, and we could all very well do with some real friggin answers! I don't have any such special insight at all. All I found out was that everyone was so obsessed with using the latest and greatest ML techniques, that there was practically no first principles thinking. At the end, everyone just sort of got too drained and gave up, which is reflected by a single participant winning pretty much the entire second round of 7-8 task prizes by the virtue of being the last man standing :-)
I have realized two things.
1) ML, at least when it comes to understanding text, is really overhyped
2) Nearly everyone who works in ML research is probably overpaid by a factor of 100 (just pulling some number out of my you know what), given that the results they have actually produced have fallen so short precisely when they were so desperately needed
and then you read this
It's like calling Russia a loser in Cold War. Technically the effect is reached; practically the side which "lost" gained possibly largest benefits.
There are no fundamental challenges remaining for level-5 autonomy. There are many small problems. And then there's the challenge of solving all those small problems and then putting the whole system together. [0]
I feel like this year is going to be another year in which the proponents of brute-force AI like Elon and Sutton will learn a bitter lesson.
[0] https://twitter.com/yicaichina/status/1281149226659901441
I hate appeals to authority as much as anybody else on HN, but I'm not sure that we could say Rich Sutton[1] is "misinformed". He's an established expert in the field, and if we discount his academic credentials then at least consider he's understandably biased towards this line of thinking as one of the early pioneers of reinforcement learning techniques[2] and currently a research scientist at DeepMind leading their office in Alberta, Canada.
[1] https://en.wikipedia.org/wiki/Richard_S._Sutton
[2] http://incompleteideas.net/papers/sutton-88-with-erratum.pdf
DNNs today can generate images that are hard to distinguish from real photos, super natural voices and surprisingly good text. They can beat us at all board games and most video games. They can write music and poetry better than the average human. Probably also drive better than an average human. Why worry about 'no progress for 50 years' at this point?
I'm not an idiot. I understand that we won't have general purpose thinking machines any time soon. But to give up entirely looking into that kind of thing, seems to me to be a mistake. To rebrand the entire field as calculating results to given problems and behaviors using existing mathematical tools, seems to do a disservice to the entire concept and future of artificial intelligence.
Imagine if the field of mathematics were stumped for a while, so investigators decided to just add up things faster and faster, and call that Mathematics.
To begin with, because they do work and much better than the new approaches in a range of domains. For example, classical planners, automated theorem provers and SAT solvers are still state-of-the-art for their respective problem domains. Statistical techniques can not do any of those things very well, if at all.
Further, because the newer techniques have proven to also be brittle in their own way. Older techniques were "brittle in the sense that they didn't deal with uncertainty very well. Modern techniques are "brittle" because they are incapable of extrapolating from their training data. For example see the "elephant in the room" paper [1] or anything about adversarial examples regarding the brittleness of computer vision (probably the biggest success in modern statistical machine learning).
Finally, AI as a field did not rely on "understanding based approaches for 50 years"; there is no formal definition of "understanding" in the context of AI. A large part of Good, Old-Fashioned AI studied reasoning, which is to say, inference over rules expressed in a logic language, e.g. this was the approach exemplified by expert systems. Another large avenue of research was that on knowledge representation. And of course, machine learning itself was part of the field from its very early days, having been named by Arthur Samuel in 1959. Neural networks themselves are positively ancient: the "artifical neuron" was first described in 1938, by Pitts & McCulloch, many years before "artificial intelligence" was even coined by John McCarthy (and at the time it was a propositional-logic based circuit and nothing to do with gradient optimisation).
In general, all those obsolete dinosaurs of GOFAI could do things that modern systems cannot - for instance, deep neural nets are unrivalled classifiers but cannot do reasoning. Conversely, logic-based AI of the '70s and '80s excelled in formal reasoning. It seems that we have "progressed" by throwing out all the progress of earlier times.
____________
[1] https://arxiv.org/abs/1808.03305
P.S. Image, speech and text generation are cute, but a very poor measure for the progress of the field. There are not even good metrics for them so even saying that deep neural nets can "generate surprisingly good text" doesn't really say anything. What is "surprisingly good text"? Surprising, for whom? Good, according to what? etc. GOFAI folk were often accused of wastig time with "toy" problems, but what exactly is text generation if not a "toy problem" and a total waste of time?
https://openai.com/blog/ai-and-compute/
The amount of compute required for Imagenet classification has been exponentially decreasing:
True, GPT-2 and -3, RoBERTa, T5 etc. are all increasingly data- and compute-hungry. That's the 'tick' your second article mentions.
We simultaneously have people doing research in the 'tock' - reducing the compute needed. ICLR 2020 was full of alternative training schema that required less compute for similar performance (e.g. ELECTRA[2]). Model distillation is another interesting idea that reduces the amount of inference-time compute needed.
So the trend isn't changing we still need bigger models to make progress in NLP and CV, while the algorithmic effeciencies are promising but they aren't giving anywhere near the same improvements as larger models.
I'm curious how long this trend will continue and if there's anything promising that can reverse this trend
Moore's law is technically "the number of transistors per unit area doubles every 24 months" [1]. The more important law is that the cost of transistors halves every 18-24 months.
That is, Moore's law talks about how many transistors we can pack into a unit area. The deeper issue is how much it costs. If we can only pack in a certain amount transistors per area but the cost drops exponentially, we still see massive gains.
There's also Wright's law that comes into play [3] that talks about dropping exponential costs just from institutional knowledge (2x in production leads to (.75-.9)x in cost).
[1] https://en.wikipedia.org/wiki/Moore%27s_law
But as mentioned in the comments below ai model training is increasing exponentially (compute required to train models has been doubling every 3.6 months) so it still far outstrips the cost savings.
They're phenomena. They're patterns we observe, and that's it. The pattern may change anytime, and that's something that should be expected. The causes may be known or unknown, but to call it a law may even make it hold true for longer, for "psychological" reasons. The law of gravity isn't influenced by what SpaceX investors think about it.
I actually wonder if having specialized AI hardware isn't the same problem as having specialized AI models, that is in the short term it will improve efficiency but in the long run prevent discovery of newer general learning strategies because they won't run faster in existing specialized hardware.
Sure it is in it's infancy but assuming that the research continues to prove that quantum computing is viable I expect it to be an even bigger deal than the move from vacuum tubes to transistors. At that point we'll be dealing with an entirely different world in computing.
On the other end of the spectrum, data and compute ARE limited and for some tasks we're at a point where the model eats up all the humanity's written works and a couple million dollars in compute and further progress has to come from elsewhere because even large companies won't spend billions of dollars in compute and humanity will not suddenly write ten times more blog articles.
And the nice thing about these large models is that you can reuse them with little fine-tuning for all sorts of other tasks. So the industry and any hacker can benefit from these uber-models without having to retrain from scratch. Of course, if they even fit the hardware available, otherwise they have to make due with a slightly lower performance.
Of course a good speech recognition system needs to model all the relevant characteristics of the human vocal tract as such, and of the many different vocal tracts of individual humans!
But this is substantially different from the notion of integrating a human-made model of the human vocal tract.
In this case the bitter lesson (which, as far as I understand, does apply to vocal tract modeling - I don't personally work on speech recognition but colleagues a few doors down do) is that if you start with some data about human voice and biology; you develop some explicit model M, and then integrate it into your system, then it does not work as well if you properly design a system that will learn speech recognition on the whole, learning an implicit model M' of the relevant properties of the vocal tract (and the distribution of these properties in different vocal tracts) as a byproduct of that, given sufficient data.
A hypothesis (which does need more research to be demonstrated, though, but we have some empirical evidence for similar things in most aspects of NLP) on the reason for this is that the human-made model M can't be as good as the learned model because it's restricted by the need to be understandable by humans. It's simplified and regularized and limited in size so that it can be reasonably developed, described, analyzed and discussed by humans - but there's no reason to suppose that the ideal model that would perfectly match reality is simple enough for that; it may well be reducible to a parameteric function that simply has too many parameters to be neatly summarizable to a human-understandable size without simplifying in ways that cost accuracy.
Human superiority doesn't come just from the brain, it comes from the environment this brain has access to - other humans, culture, tools, nature, and the bodily affordances (hands, feet, eyes, ability to assimilate organic food...). AI needs a body and an environment to evolve in.
Gosh, what an awkward pose, that first picture.
And certainly fine-tuning and distillation are part of the story why we wanted these large do-all-be-all models in the first place, but the question of what's next for the state of the art - and that currently would be featurization through a large transformer model (i.e. BERT, ERNIE, GPT-2) with some deep-but-not-huge task-specific model on top - isn't simply answered by "more compute".
As long as our proof of concept solutions don't yet solve the task appropriately, as long as the solution is weak and/or brittle and worse than what we need for the main partical applications, most of the research focus - and the research progress - will be on models that try and give better results. It makes sense to disregarding the compute cost and other impractical inconveniences when working on pushing the bleeding edge, trying to make the previously impossible things possible
However, when tasks are "solved" from the academic proof-of-concept perspective, then generally the practical, applied work on model efficiency can get huge reductions in computing power required. But that happens elsewhere.
The concept of technology readiness level (https://en.wikipedia.org/wiki/Technology_readiness_level) is relevant. For the NLP and CV technologies that are in TRL 3 or 4, the efficiency does not really matter as long as it fits in whatever computing clusters you can afford; this is mainly an issue for the widespread adoption of some tech in industry by the time the same tech is in TRL 6 or so, and this work mostly gets done by different people in different organizations with different funding sources than the initial TRL 3 research.
The question is why you say that humans perform an unconscious search when they play Go. And what kind of search is it, other than unconscious? Could you describe it, e.g. in algorithmic notation? I mean, I'm sure you couldn't because if you could then the problem of teaching a computer to play Go as well as a human would have been solved years and years ago. But, if you can't describe what you're doing, then how do you know it's a "search"?
Note that in AI, when we talk of "search" (edit: at least, in the context of game-playing) we mean something very specific: an algorithm that examines the nodes of a tree and applies some criterion to label each examined node as a target node or not a target node. Humans are absolutely awful at executing such an algorithm with our minds for any but the most trivial of trees, at least compared to computers.
It's really fun to watch his commentary because he relentlessly plays "variations" — possible next moves and sequences — while waiting for the players, explaining the tradeoffs between moves and the consequences they lead to a few steps ahead in the game.
I don't know what to call "variations" but a tree search with heuristics. He does it slowly to explain it to the audience, but I have no doubt the same process runs much faster in his mind.
If you can avoid making "obvious" (in retrospect) mistakes, you will be a competitive chess or go player.
And given that we have millions of neurons to suppress activity in one node if activity was high enough in another node, I would summarize the brain as effectively a huge chain of look up tables.
To clarify, you can come up with an explanation of anything that you do, or observe yourself or another person do. For example, you might explain how you hit a ball with racket in tennis or with a bat in baseball, etc, but that doesn't mean that the process you are describing is the process that your mind (let alone your brain) actually follows.
If nothing else because such a description will necessarily fudge important steps. For example, if I describe myself walking as "I put one foot in front of the other" - have I explained enough about walking that it can now be reproduced mechanically? Experience teaches that -no.
I don't understand why you keep making personal comments like that about me. I suspect you don't realise that they are unpleasant. Please let me make it clear: such personal comments are unpleasant. Could you please stop them? Thank you.
In this work we present the MuZero algorithm which, by combining _a tree-based search_ with a learned model, achieves superhuman performance in a range of challenging and visually complex domains, withoutany knowledge of their underlying dynamics.
(My underlining)
If I remember correctly, MuZero is model-free in the sense that it learns its own evaluation function and reward policy etc (also going by the abstract). But it retains MCTS.
Indeed, it wouldn't really make sense to drop MCTS from the architecture of a system designed to play games. I mean, it would be really hard to justify discarding a component that is well known to work and work well, both from an engineering and a scientific point of view.
_________________
I quote from the arxiv paper again:
Appendix B Search
We now describe the search algorithm used by MuZero. Our approach is based upon Monte-Carlo tree search with upper confidence bounds, an approach to planning that converges asymptotically to the optimal policy in single agent domains and to the minimax value function in zero sum games [22].
So I'm sorry but I really don't understand what you mean here.
Also, as discussed in other comments, it's dangerous to assume anyhing about how "a human does" anything to do with any kind of mental calculation. Whatever MuZero does and however good or bad it does it, there is nothing to tell us that it does it as a human does.