Troels Petersen - Using Graph Neural Networks for classification and reconstruction in IceCube
→
Europe/Berlin
SR5 (Villa)
SR5 (Villa)
Description
Basic Machine Learning algorithms have long been used in particle physics, especially lately, and for good reasons. They often manage to get more information out of the raw data and with a shorter inference time. However, this is not without challenges. First of all, algorithms are typically trained on simulated data, which does not perfectly represent the true data, leading to suboptimal performance, but also the basic ML algorithms - boosted decision trees (BDTs) and neural networks (NNs) - do not apply well to some cases, such as the IceCube detector.
We have been working on applying more advanced ML methods (GNNs) to IceCube data, trying to explore the possibilities of getting the best out of our fantastic detectors. I’ll talk about the method and the results gotten so far.
