arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2009.03192cs.LGnucl-thphysics.comp-phstat.ML
收藏

基于Born级数的神经网络扰动理论

A Neural Network Perturbation Theory Based on the Born Series

Helmholtz-Institut für Strahlen- und Kernphysik(亥姆霍兹辐射与核物理研究所);Bethe Center for Theoretical Physics(贝特理论物理中心);Universität Bonn(波恩大学);Institute for Advanced Simulation(高级仿真研究所);Institut für Kernphysik(核物理研究所);Jülich Center for Hadron Physics(于利希强子物理中心);Forschungszentrum Jülich(于利希研究中心);Tbilisi State University(第比利斯国立大学)

Bastian Kaspschak, Ulf-G. Meißner

摘要:

在过去的十年中,使用epsilon深度神经网络(DNN)的深度学习已经成为解决各种基于数据的理论物理问题的一种有吸引力的方法。有一个明显的趋势是更深入的架构包含越来越强大和涉及的层。相反,DNN的泰勒系数仍然主要出现在可解释性研究中,其中它们最多计算到一阶。然而,特别是在理论物理学中,许多问题也受益于访问高阶。这种差距激发了神经网络(NN)泰勒展开的一般公式。限制我们的分析,多层感知器(MLP)和引入数量,我们称为传播和顶点,都取决于MLP的权重和偏见,我们建立了一个图论的方法。类似于量子场论中的费曼规则,我们可以系统地将包含传播子和顶点的图分配给相应的偏导数。检查这种方法的S-波散射长度的浅电位,我们观察到神经网络,以适应他们的衍生物主要是目标函数的泰勒展开的领导阶。为了避免这个问题,我们提出了一个迭代神经网络扰动理论。在每次迭代中,我们消除了前导顺序,这样下一个前导顺序可以在后续迭代中忠实地学习。在进行两次迭代后,我们发现一阶和二阶玻恩项在各自的迭代过程中得到了正确的调整。最后,我们结合这两个结果,找到一个代理,作为机器学习的二阶玻恩近似。

英文摘要:

Deep Learning using the eponymous deep neural networks (DNNs) has become an attractive approach towards various data-based problems of theoretical physics in the past decade. There has been a clear trend to deeper architectures containing increasingly more powerful and involved layers. Contrarily, Taylor coefficients of DNNs still appear mainly in the light of interpretability studies, where they are computed at most to first order. However, especially in theoretical physics numerous problems benefit from accessing higher orders, as well. This gap motivates a general formulation of neural network (NN) Taylor expansions. Restricting our analysis to multilayer perceptrons (MLPs) and introducing quantities we refer to as propagators and vertices, both depending on the MLP's weights and biases, we establish a graph-theoretical approach. Similarly to Feynman rules in quantum field theories, we can systematically assign diagrams containing propagators and vertices to the corresponding partial derivative. Examining this approach for S-wave scattering lengths of shallow potentials, we observe NNs to adapt their derivatives mainly to the leading order of the target function's Taylor expansion. To circumvent this problem, we propose an iterative NN perturbation theory. During each iteration we eliminate the leading order, such that the next-to-leading order can be faithfully learned during the subsequent iteration. After performing two iterations, we find that the first- and second-order Born terms are correctly adapted during the respective iterations. Finally, we combine both results to find a proxy that acts as a machine-learned second-order Born approximation.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑