4
$\begingroup$

Apart from being a student of Physics I'm also a developer and recently in the context of cloud development I've been hearing a lot about cloud based Machine Learning. After looking at some examples of it I found it quite impressive.

With Machine Learning we can process lots of data and we can have the computer analyze it in many ways. As I saw we can use this to detect patterns, find anomalies in the data and so forth.

Although I'm much more interested in theoretical Physics and I don't understand much about experimental Physics I believe that as experiments in Physics gets more complex the analysis of the gathered data also increases in complexity.

This made me wondering if those methods of cloud based machine learning can be used in Physics to analyse experimental data. Is this something that's been used recently or it has no use at all? If it has no use at all, why is that the case?

$\endgroup$
  • 2
    $\begingroup$ I know the collider guys have flirted with it a time or three, and I think I heard that they've been at it again. Beyond that I couldn't say. $\endgroup$ – dmckee Feb 23 '16 at 1:12
  • 2
    $\begingroup$ The main problem with machine learning is that it's usually very hard to understand the biases that the machine has learned (which depend on the training sets). Sampling and analysis bias are extremely dangerous in science because they can mimic effects that are just not there. I would be very careful to use these techniques without extensive controls... which, of course, makes their use far less attractive than the "fire and forget" mentality that has created an appetite for them. $\endgroup$ – CuriousOne Feb 23 '16 at 1:13
  • $\begingroup$ @dmckee: I can confirm that. Neural networks were hugely popular among many research groups some twenty years ago. Part of that was driven by grants... one could get money for track finding and other algorithm research by reclassifying it as applications of machine learning and AI. I have not seen these algorithms perform significantly better than more direct approaches like ordinary fitters, Bayesian estimators etc.. With sufficient "massaging" one can come up with many ways to skin the data cat that are all more or less equally good or bad, depending on the metrics. $\endgroup$ – CuriousOne Feb 23 '16 at 1:16
  • $\begingroup$ @CuriousOne Machine learning techniques typically involve a "training" data set, which has been processed by either a human or some other trusted algorithm. What you call "extensive controls" are, I think, a typical part of the machine learning procedure. Do you have a reference for the notion that it's hard to understand the biases? I'm genuinely interested (since I work in a machine learning lab and know almost nothing about it). $\endgroup$ – DanielSank Feb 23 '16 at 2:09
  • $\begingroup$ @DanielSank: The last experiment that I was involved in that had a machine learning algorithm was willing to sacrifice half their real data set for the learning process at some point. That was an insane idea... and I don't think they did, in the end, but some of the folks in the data analysis group went totally blind to more rational ways to analyze that data. The religious belief that the machine would be unbiased to the real data if it was trained on the real data had taken a foothold. ML is as good or as bad as we make it... it can never be better than that. :-) $\endgroup$ – CuriousOne Feb 23 '16 at 2:13
2
$\begingroup$

Machine learning is particularly useful in situations where the underlying model connecting input and output is either very complicated, or not understood, or both. Another situation in which machine learning can be useful is if the quantity of data is greater than the researcher (or team) can look at directly. Neither of these circumstances is very palatable to the typical physicist. Nearly all of what we would call physics consists of fairly simple models, and the urge to construct and understand simple models drives much of what we do. And, as far as the second point is concerned, collecting data has costs associated with it, and so most people don't want to collect more data than they will be able to directly use.

There are two major exceptions to this: observational astronomy (and in particular, big survey projects like SDSS) and high energy experiments like LHC. In both of these cases, machine learning has been used to good effect. (For example, see the Kaggle competitions for detecting the Higgs and classifying galaxies.)

Another promising application is in turbulence modeling (currently under development by Karthik Duraisamy and his group at the University of Michigan). Subgridscale turbulence models are already statistical, so it makes sense to use statistical learning methods to refine them.

Personally, I agree with you that machine learning ought to have a place in the physicists toolbox. We are, in general, comfortable with statistical descriptions of nature; we make heavy use of approximation methods; we would like to think we know how to make careful measurements, handle data, and propagate and account for uncertainties; and we have long been aware that there are situations in which a computational approach is the best (if not only) approach to a problem.

$\endgroup$
  • 3
    $\begingroup$ I beg to differ that most of physics has only small amounts of data and simple models. That hasn't been true since, at least, WW II. Physicists were consistently at the forefront of computing and by the 1970s an experimental physics lab without a local PDP was considered a place to avoid. Today access to a supercomputing facility (and if it's nothing but a bunch of PCs with graphics cards) is almost a must, even for relatively small experiments. GByte size data sets are the new normal... and why would they not be? A GByte can be churned trough an ordinary PC in minutes. $\endgroup$ – CuriousOne Feb 23 '16 at 2:03

Not the answer you're looking for? Browse other questions tagged or ask your own question.