6 ms·

I am studying physics now (with a CS and math background) and I feel obligated to get up to date on AI and to develop a good working philosophy of how it can be

I am studying physics now (with a CS and math background) and I feel obligated to get up to date on AI and to develop a good working philosophy of how it can be meaningfully used in scientific work. Not that all "apply machine learning to X" approaches aren't interesting, but I lack enough understanding to know whether these are popping up everywhere because people feel obligated to apply new methods. For example, the Fourier transform is deep and interesting and there are libraries and standards and ways to transform different objects over clusters, etc., but I wouldn't say good scientific research is about finding a place to apply a Fourier transform (maybe :)). I am new to this though.

dotnet00 3y ago
I'm nearing the end of my PhD in computer engineering, but because my uni is associated with a big physics research lab, my work ended up being pretty involved with physics. I've had to go through a similar learning experience regarding how to meaningfully apply AI (and computation in general) to perform meaningful science. I agree that simply applying a fourier transform isn't meaningful research, but there are fields like fourier optics, where you're effectively just doing that, with the goal of modeling wave optics. The kinds of issues I had in mind wrt poor uses of AI are things like inverse problem solving methods which train a model as a black box and neglect to involve physics informed feedback, so the results are of questionable value, or hyperparameter estimators which need an unrealistic amount of data or produce blatantly unrealistic estimates because they're naive and the AI "researcher" has no interest in actually understanding the physical logic to the parameters. My research hasn't been with AI, just with simulations of certain physical systems, and one thing that was constantly drilled into me by the physicists was to be really careful about how far I stretch computational techniques, because ultimately none of this digital stuff means much if it becomes detached from reality. It can even cause actual damage because sometimes expensive purchases will be made on the basis of simulations.
- timeforcomputer 3y ago
  Thanks for the pointer to Fourier optics! I mentioned Fourier because it is something I still don't "get" even after studying and using these methods. I am still learning, but I have so far only applied inverse methods to systems where the physics is well-defined and the model already accepted as "real". For example, CT scanner algorithms which model the photon counters and attenuation through matter (although this can get complicated...), with the output being data in well-defined format (like the distribution of matter). I do see people using inverse solvers or AI to derive something which isn't an "image" but a model of the physics (does that make sense?). The extreme version of this which I have seen once before, is a paper using AI to extract algebraic forms of PDEs from videos of fluids. To what extent does the AI user have to understand/"pre-model" the physics, and to what extent does the AI generate understanding in some way? Apologies if this is misinformed as I am still outside the research community, and don't very actively read papers. Currently I am trying to learn background to understand electromagnetic properties of materials with an interest to apply this to computer graphics. I eventually want to learn what processes are currently done for extracting material properties from scans, etc., along the lines of the classic gonioreflectometers. In this space, from the papers I look at, it looks like AI is unavoidable and I am sure an extremely useful tool but I definitely want to be careful not to get lost in misunderstanding.
  - dotnet00 3y ago
    >To what extent does the AI user have to understand/"pre-model" the physics, and to what extent does the AI generate understanding in some way? The conclusion that I reached is that you need to understand at least enough to be able to handle basic scrutiny from people in the same field. So, for example, if you come up with an AI for deriving some model of the physics of a CT scanner, you should at least understand the problem you're tackling enough to make meaningful comparisons to any popular competing approaches that might exist. That understanding should automatically encompass things like understanding how your variables are bounded/how they behave. A contrived example could be with an AI intended to guess 2 prime factors of a given number. Imagine someone making such a tool without even knowing what a prime number is, where the only benchmark they use for correctness is if its outputs fit the dataset of products of prime numbers given. The tool would be clearly useless, because while to the AI researcher it'd seem fine that it outputs 36, which is close to the intended 37, to anyone who knows about prime numbers it'd be an immediate warning sign needing further elaboration and explanation. The researcher could protest that as an AI expert, they don't know anything about prime numbers, but that doesn't really matter in research. Similarly, due to my non-physics background, while I was (and to an extent still am) given significant leeway in terms of my ability to respond to detailed physics questions about my work, by now I am expected to be able to handle questions about things that are common knowledge among physicists in this field (so, for example, knowing what performing a Fourier transform means in our context, understanding the high level functioning of the key components of the machines we work with, having some intuition for when a simulation result is unrealistic etc). If I were to talk about how I've simulated some of these components, but drew a complete blank when asked how my work improved upon the common knowledge means of simulating them, I'd obviously look like a complete idiot :P It may seem like a pretty low and obvious bar, but surprisingly I have seen many cases of researchers who didn't even care about reaching that level.

dotnet00 3y ago

I'm nearing the end of my PhD in computer engineering, but because my uni is associated with a big physics research lab, my work ended up being pretty involved with physics. I've had to go through a similar learning experience regarding how to meaningfully apply AI (and computation in general) to perform meaningful science. I agree that simply applying a fourier transform isn't meaningful research, but there are fields like fourier optics, where you're effectively just doing that, with the goal of modeling wave optics. The kinds of issues I had in mind wrt poor uses of AI are things like inverse problem solving methods which train a model as a black box and neglect to involve physics informed feedback, so the results are of questionable value, or hyperparameter estimators which need an unrealistic amount of data or produce blatantly unrealistic estimates because they're naive and the AI "researcher" has no interest in actually understanding the physical logic to the parameters. My research hasn't been with AI, just with simulations of certain physical systems, and one thing that was constantly drilled into me by the physicists was to be really careful about how far I stretch computational techniques, because ultimately none of this digital stuff means much if it becomes detached from reality. It can even cause actual damage because sometimes expensive purchases will be made on the basis of simulations.

timeforcomputer 3y ago
Thanks for the pointer to Fourier optics! I mentioned Fourier because it is something I still don't "get" even after studying and using these methods. I am still learning, but I have so far only applied inverse methods to systems where the physics is well-defined and the model already accepted as "real". For example, CT scanner algorithms which model the photon counters and attenuation through matter (although this can get complicated...), with the output being data in well-defined format (like the distribution of matter). I do see people using inverse solvers or AI to derive something which isn't an "image" but a model of the physics (does that make sense?). The extreme version of this which I have seen once before, is a paper using AI to extract algebraic forms of PDEs from videos of fluids. To what extent does the AI user have to understand/"pre-model" the physics, and to what extent does the AI generate understanding in some way? Apologies if this is misinformed as I am still outside the research community, and don't very actively read papers. Currently I am trying to learn background to understand electromagnetic properties of materials with an interest to apply this to computer graphics. I eventually want to learn what processes are currently done for extracting material properties from scans, etc., along the lines of the classic gonioreflectometers. In this space, from the papers I look at, it looks like AI is unavoidable and I am sure an extremely useful tool but I definitely want to be careful not to get lost in misunderstanding.
- dotnet00 3y ago
  >To what extent does the AI user have to understand/"pre-model" the physics, and to what extent does the AI generate understanding in some way? The conclusion that I reached is that you need to understand at least enough to be able to handle basic scrutiny from people in the same field. So, for example, if you come up with an AI for deriving some model of the physics of a CT scanner, you should at least understand the problem you're tackling enough to make meaningful comparisons to any popular competing approaches that might exist. That understanding should automatically encompass things like understanding how your variables are bounded/how they behave. A contrived example could be with an AI intended to guess 2 prime factors of a given number. Imagine someone making such a tool without even knowing what a prime number is, where the only benchmark they use for correctness is if its outputs fit the dataset of products of prime numbers given. The tool would be clearly useless, because while to the AI researcher it'd seem fine that it outputs 36, which is close to the intended 37, to anyone who knows about prime numbers it'd be an immediate warning sign needing further elaboration and explanation. The researcher could protest that as an AI expert, they don't know anything about prime numbers, but that doesn't really matter in research. Similarly, due to my non-physics background, while I was (and to an extent still am) given significant leeway in terms of my ability to respond to detailed physics questions about my work, by now I am expected to be able to handle questions about things that are common knowledge among physicists in this field (so, for example, knowing what performing a Fourier transform means in our context, understanding the high level functioning of the key components of the machines we work with, having some intuition for when a simulation result is unrealistic etc). If I were to talk about how I've simulated some of these components, but drew a complete blank when asked how my work improved upon the common knowledge means of simulating them, I'd obviously look like a complete idiot :P It may seem like a pretty low and obvious bar, but surprisingly I have seen many cases of researchers who didn't even care about reaching that level.