3 ms·

What in the world Most "diffusion models" use similar VAE to transformer backbone architectures. Diffusion isn't an architecture, it's a problem framing As fo

by advael 2y ago

What in the world Most "diffusion models" use similar VAE to transformer backbone architectures. Diffusion isn't an architecture, it's a problem framing As for the rest of this, I'm torn between liking the poetry of it and pointing out that this is kind of that thing where you say something like it's supposed to be a mind-blowing insight when it's well-known and pretty obvious. Most people familiar with learning theory already understand learning algorithms of any kind as a subset of probabilistic search algorithms with properties that make them responsive to data. The idea that the structure of the information processing system doesn't matter and there's just this general factor of learning capacity a thing has is... not well supported by the way in which research has progressed in the entire period of time when this has been relevant to most people? Sure, in theory any neural network is a general function approximator and could in theory learn any function it's complex enough to represent. Also, we can arrive at the solution to any computable problem by representing it as a number and guessing random numbers until we can verify a solution. Learning algorithms can almost be defined as attempts to do better search via structured empiricism than can be done with the assumption that structure doesn't matter. Like, sometimes multiple things work, sure. That doesn't mean it's arbitrary TL;DR: Of course learning is a kind of search, but discovering structures that are good at learning is the whole game

Xcelerate 2y ago
Yeah, I really don’t understand this recently popular viewpoint that the algorithm doesn’t matter, just how much data you throw at it. It doesn’t seem to be based on anything more than wishful thinking. One can apply Hutter search to solve just about any problem conceivable given the data and guess what—you’ll approach the optimal solution! The only downside is that this process will take more time than available in our physical universe. I think people forget the time factor and how the entire field of computational complexity theory arose because the meta problem is not that we can’t solve the problem—it’s that we can’t solve it quickly enough on a timescale that matters to humans. Current NN architectures are missing something very fundamental related to the efficiency of problem solving, and I really don’t see how throwing more data at them is going to magically convert an EXPTIME algorithm into a PTIME one. (I’m not saying NNs are EXPTIME; I’m saying that they are incapable of solving entire classes of problems that have both PTIME and EXPTIME solutions, as the NN architecture is not able to “discover” PTIME solutions, thus rendering them incapable of solving those classes of problems in any practical sense).
- advael 2y ago
  Also, one of the major classes of problem that gets solved and we view as "progress" in machine learning is framing problems. Like we couldn't define "draw a good picture" in a way we could actually assess well, GANs and Diffusion turn out to be good ways to phrase problems like that. In the former case, it creates a way to define the problem as "make something indistinguishable from an example pulled from this dataset" and in the latter case, "I've randomized some of these pixels, undo that based on the description" The idea of "efficiency" and "progress" is this post-hoc rationalization that people who never understood the problem, pay people to understand the problem, apply to problems once they have a working solution in hand. It's a model that is inherently as dumb as a model can be, and the assumption it makes is that there is some general factor of progress on hard problems that can be dialed up and down. Sure, you can pay more scientists and probabilistically increase the rate at which problems are solved, but you can't predict how long it will take, how much failure it will involve, whether a particular scientist will solve a particular problem at all, whether that problem is even solvable in principle sometimes. Businesspeople and investors like models where you put in money and you get out growth at a predictable rate with a controllable timescale, and if this doesn't work you just kick it harder, and this ill fits most frontier research. Hell, it ill suits a lot of regular work.
  - visarga 2y ago
    > Sure, you can pay more scientists and probabilistically increase the rate at which problems are solved, but you can't predict how long it will take, how much failure it will involve, whether a particular scientist will solve a particular problem at all, whether that problem is even solvable in principle sometimes. Fully agree, search is hard, unpredictable and expensive. Also a matter of luck, being at the right place and time, and observing something novel. That is why I put the emphasis of AI doing search, not just imitating humans.
    - advael 2y ago
      Okay, but what does that mean? AI is a search process. Do you mean you want the AI to formulate queries? Test hypotheses? Okay. How? What does that mean? What we know how to do is to mathematically define objective functions and tune against them. What objective function describes the behavior you want? Is there some structural paradigm we can use for this other than tuning the parameters on a neural network through optimization toward an objective function? If so, what is it? I'm sorry to be a little testy but what you've basically said is "We should go solve the hard problems in AI research". Dope. As an AI researcher I fully agree. Thanks. Am I supposed to clap or something?
visarga 2y ago
> Of course learning is a kind of search, but discovering structures that are good at learning is the whole game No, you missed the essential. I mentioned search in the context of discovery, or in other words expanding knowledge. Training neural nets is also a search for the best parameters that fit the data, but it's secondary. Many architectures work, there have been a thousand variations for the transformer architectures and plenty of RNN-like approaches since 2017 when transformer was invented, and none of them is better than the current one or significantly worse. Also, considering human population, the number of neurons in the brain, synapses and wiring are very different at micro level from person to person, yet we all learn. The difference between the top 5% and bottom 5% humans is small compared with other species, for example. What makes a big difference between people is education, in other words experiences, or training data. To return to the original idea - AI that simply learns to imitate human text is capable only of remixing ideas. But an AI that actively explores can discover novel ideas, like AlphaZero and AlphaTensor. In both these cases search played a major role. So I was generalizing the concept of "search" across many levels of optimization, from protein folding to DNA and human intelligence. Search is essential for progress across the stack. Even network architecture evolves by search - with human researchers.

by advael 2y ago

Xcelerate 2y ago

Yeah, I really don’t understand this recently popular viewpoint that the algorithm doesn’t matter, just how much data you throw at it. It doesn’t seem to be based on anything more than wishful thinking. One can apply Hutter search to solve just about any problem conceivable given the data and guess what—you’ll approach the optimal solution! The only downside is that this process will take more time than available in our physical universe. I think people forget the time factor and how the entire field of computational complexity theory arose because the meta problem is not that we can’t solve the problem—it’s that we can’t solve it quickly enough on a timescale that matters to humans. Current NN architectures are missing something very fundamental related to the efficiency of problem solving, and I really don’t see how throwing more data at them is going to magically convert an EXPTIME algorithm into a PTIME one. (I’m not saying NNs are EXPTIME; I’m saying that they are incapable of solving entire classes of problems that have both PTIME and EXPTIME solutions, as the NN architecture is not able to “discover” PTIME solutions, thus rendering them incapable of solving those classes of problems in any practical sense).

advael 2y ago
Also, one of the major classes of problem that gets solved and we view as "progress" in machine learning is framing problems. Like we couldn't define "draw a good picture" in a way we could actually assess well, GANs and Diffusion turn out to be good ways to phrase problems like that. In the former case, it creates a way to define the problem as "make something indistinguishable from an example pulled from this dataset" and in the latter case, "I've randomized some of these pixels, undo that based on the description" The idea of "efficiency" and "progress" is this post-hoc rationalization that people who never understood the problem, pay people to understand the problem, apply to problems once they have a working solution in hand. It's a model that is inherently as dumb as a model can be, and the assumption it makes is that there is some general factor of progress on hard problems that can be dialed up and down. Sure, you can pay more scientists and probabilistically increase the rate at which problems are solved, but you can't predict how long it will take, how much failure it will involve, whether a particular scientist will solve a particular problem at all, whether that problem is even solvable in principle sometimes. Businesspeople and investors like models where you put in money and you get out growth at a predictable rate with a controllable timescale, and if this doesn't work you just kick it harder, and this ill fits most frontier research. Hell, it ill suits a lot of regular work.
- visarga 2y ago
  > Sure, you can pay more scientists and probabilistically increase the rate at which problems are solved, but you can't predict how long it will take, how much failure it will involve, whether a particular scientist will solve a particular problem at all, whether that problem is even solvable in principle sometimes. Fully agree, search is hard, unpredictable and expensive. Also a matter of luck, being at the right place and time, and observing something novel. That is why I put the emphasis of AI doing search, not just imitating humans.
  - advael 2y ago
    Okay, but what does that mean? AI is a search process. Do you mean you want the AI to formulate queries? Test hypotheses? Okay. How? What does that mean? What we know how to do is to mathematically define objective functions and tune against them. What objective function describes the behavior you want? Is there some structural paradigm we can use for this other than tuning the parameters on a neural network through optimization toward an objective function? If so, what is it? I'm sorry to be a little testy but what you've basically said is "We should go solve the hard problems in AI research". Dope. As an AI researcher I fully agree. Thanks. Am I supposed to clap or something?

visarga 2y ago

> Of course learning is a kind of search, but discovering structures that are good at learning is the whole game No, you missed the essential. I mentioned search in the context of discovery, or in other words expanding knowledge. Training neural nets is also a search for the best parameters that fit the data, but it's secondary. Many architectures work, there have been a thousand variations for the transformer architectures and plenty of RNN-like approaches since 2017 when transformer was invented, and none of them is better than the current one or significantly worse. Also, considering human population, the number of neurons in the brain, synapses and wiring are very different at micro level from person to person, yet we all learn. The difference between the top 5% and bottom 5% humans is small compared with other species, for example. What makes a big difference between people is education, in other words experiences, or training data. To return to the original idea - AI that simply learns to imitate human text is capable only of remixing ideas. But an AI that actively explores can discover novel ideas, like AlphaZero and AlphaTensor. In both these cases search played a major role. So I was generalizing the concept of "search" across many levels of optimization, from protein folding to DNA and human intelligence. Search is essential for progress across the stack. Even network architecture evolves by search - with human researchers.