Cool! When was that?
To say the truth, I would like to use this to develop new techniques for looking for new phenomena at particle colliders. I have only vague ideas for the moment, and the rest requires time that I don't have so far :( But I will try to slowly start moving along this path ^^
Dug up an old write for you. Before 2010 lol :) - https://steemit.com/technology/@kevinwong/i-played-a-part-in-the-formation-of-the-surveillance-state
What kind of data are you looking to process?
In a collisions at the LHC, we get final state products that consist of many particles. We then cluster them in some way (that contains some free parameters and methods) and study the output of this clustering. From this output, we can study several observables, properties, etc...
Now, in terms of searches for new phenomena, we have the signal (the new phenomenon) and the background (the Standard Model expectation). We want to know what to check (how to select our collision events) to maximise the signal and reduce the background.
I would like to design something that first is capable to chose the reconstruction method automatically in the aim of improvement a signal-to-noise ratio, and second capable to chose the observable to focus on for distinguishing the signal from the background.
Dunno whether it is clear enough (I tried to be concise instead).
Thanks for sharing the other post. This dates from before my steemit time ^^ What I want to do is very different from that. It is more for the beauty of science than for making anyone rich :)
Btw referencing your earlier reply -
So machine-learning will be applied to which part of this process? Pattern recognition for collider configurations, or the output (both "live" and archived data) ?
I think this will need some lecture of its own - I'll go do some research on it. Hopefully there's something decent online :)
Mmmh I will have a look. I have never heard about them (I am also a newbie in ML).
Just to self advertise my older post. You may want to start from there ^^
I would like to use ML to work out the output (could it be a simulated collision or a real one, but it is better to start with simulated stuff since data are not publicly released yet and only experimenters have access to them) and try to reconstruct what happened.
Bookmarked for weekend reading :). Btw a lil off-tangent, have you seen Numerai? It's a pretty interesting company, sometihng to do with machine learning and data encryption for worldwide collaboration.
ML to work on pattern-recognizing the output right? If that's the case then i pretty much understood what you're trying to say..
Exactly, with the additional point that we have no clue about which pattern is better from the start. Also, varying from one model to another, the best pattern may change.
This is IMO a perfect problem for designing a ML solution.