Particle physics experiments can be regarded as sophisticated cameras, sensitive to a very special type of “light”, i.e. charged radiation (particles). This “light” originates from other particles colliding with matter constituents or other “free” particles. Ultimately, doing particle physics is taking an enormous amount of very special photos, then discarding those less interesting and saving a small fraction of them (one out of one million or so). Finally, it corresponds to focussing on certain pages of this photo album and looking for something unexpected.
Up to 2021, ATLAS and CMS at the LHC saved hundreds of billions of “pictures”, an amount of information that exceeds by more than a factor of 100 that available from previous generation experiments, like CDF and D0 at Tevatron. This explains why the expression “big data” never startled particle physicists: they grow up with data that are always “big” and they routinely apply the most sophisticated techniques to analyse them.
Nowadays, high-energy physics analyses make use of Deep Learning models at various stages: the flavour of particle jets is determined with neural networks as well as events are categorised as signal or background on the basis of the response of BDTs or other DL algorithms. Recent developments aim at implementing DL neural networks on FPGAs to perform the event selection already at trigger level.
In 2017 the “Deep learning for Particle Physics” initiative (DeepPP) started as a collaboration among the Depart of Physics of the University of Trento and the MPBA center of Fondazione Bruno Kessler. The TIFPA institute was also involved. The idea was to pursue interests in common between particle physicists (from UniTn and TIFPA) and data scientists (from FBK). In fact, particle physics can offer a wealth of well-structured and homogenous data, something that other big-data-set not always guarantee. Moreover, data-science evolves so fast that physicists often miss the latest opportunities about algorithms, models and solutions.
Since then, the DeepPP initiative has involved scientists, students and interns, doing physics within international collaborations (ATLAS, Limadou) and developing new DL algorithms for future experiments measuring particle interactions.
The most important activities of DL development have regarded:
- flavour tagging in the ATLAS experiment
- tagging of Higgs-to-bb large-radius jets against QCD-generated ones in ATLAS
- particle identification and event reconstruction in Limadou-HEPD
Now, the DeepPP initiative focusses on these problems:
- feature ranking. In short, you have M possible input categories to input your model and you want to reduce the length of the input vector to N. How to choose the best set of N variables for the specific problem you want to solve? How to make this choice robust against data set variations?
- uncertainty estimation. After being trained, models provide predictions for any input vector. But how accurate is this prediction? We know that the prediction is the best possible with the model implemented, but we are very often unaware of how distant the result could be from truth. How to estimate (and possibly reduce) the uncertainties associated with deep neural network predictions?
The DeepPP initiative makes use of a set of pseudo-data produced in Trento with a detailed simulation of particle interactions and detector response. As a consequence, results and achievements are immediately applicable to the analysis of real data from particle physics experiments, but they are devised to get generalised to other fields of application.