Exploring the challenges of using machine learning for ALICE jet measurements

Ongoing challenges for the use of machine learning (ML) for jet physics include the interpretability and the quantification of uncertainties/bias in such measurements. One use case of ML that illustrates its potential is the jet measurements in the ALICE detector discussed in this talk. In this case, ML is used in order to correct the measured jets for the large fluctuating background in heavy-ion collisions. The ML estimator utilizes properties of jets from simulation in training, including the constituents of the jet. This analysis does not utilize deep learning methods but instead utilizes a shallow neural network for increased interpretability. The bias of interest in this approach is introduced by including constituent information in training. This introduces a dependence on the fragmentation pattern in simulation, which has been shown to differ from the fragmentation in Pb--Pb data. In an effort to understand and quantify this bias, three modifications of the fragmentation utilized in training were studied, illustrating the degree to which the ML estimator is robust to the biases explored. Such results demonstrate ways in which ML approaches in jet physics can overcome the contemporary challenges of the field in order to make useful, interpretable measurements.