标题： 机器学习和深度学习技术在ALICE重离子碰撞中的应用 显示英文标题

标题： Machine and deep learning techniques in heavy-ion collisions with ALICE

摘要： 近年来，机器学习工具已被成功应用于高能物理中的各种问题。 一个典型的例子是物理对象的分类。 监督机器学习方法通过考虑可观测的相关性并从示例中学习最佳选择，例如从蒙特卡罗模拟中，从而显著提高分类问题的性能。 更令人鼓舞的是深度学习技术的应用。 像深度卷积网络这样的方法可能能够捕捉到低层次参数中的特征，而这些特征默认的基于切割的方法并未利用。 这些想法对于重离子碰撞中的测量特别有益，因为多重性非常大。 事实上，与基于切割的方法相比，机器学习方法在具有大量自由度的系统中潜在表现更好。 此外，许多关键的重离子可观测量在低横向动量下最为有趣，此时背景事件占主导地位，信号噪声比相当低。 在这项工作中，将介绍ALICE在重离子碰撞中机器学习和深度学习应用的最新进展，重点介绍一种基于深度学习的b-喷注标记方法以及低质量双电子的测量。 虽然b-喷注标记器结合了浅层全连接网络和深度卷积网络，但低质量双电子的测量使用了梯度提升和浅层神经网络。 与默认的基于切割的方法相比，这两种方法都非常有前景。 显示英文摘要

摘要： Over the last years, machine learning tools have been successfully applied to a wealth of problems in high-energy physics. A typical example is the classification of physics objects. Supervised machine learning methods allow for significant improvements in classification problems by taking into account observable correlations and by learning the optimal selection from examples, e.g. from Monte Carlo simulations. Even more promising is the usage of deep learning techniques. Methods like deep convolutional networks might be able to catch features from low-level parameters that are not exploited by default cut-based methods. These ideas could be particularly beneficial for measurements in heavy-ion collisions, because of the very large multiplicities. Indeed, machine learning methods potentially perform much better in systems with a large number of degrees of freedom compared to cut-based methods. Moreover, many key heavy-ion observables are most interesting at low transverse momentum where the underlying event is dominant and the signal-to-noise ratio is quite low. In this work, recent developments of machine- and deep learning applications in heavy-ion collisions with ALICE will be presented, with focus on a deep learning-based b-jet tagging approach and the measurement of low-mass dielectrons. While the b-jet tagger is based on a mixture of shallow fully-connected and deep convolutional networks, the low-mass dielectron measurement uses gradient boosting and shallow neural networks. Both methods are very promising compared to default cut-based methods.