物理信息随机配置机:一种用于非线性微分方程的无反向传播快速训练神经网络
文章背景与核心概要
物理信息神经网络(PINNs)彻底改变了复杂微分方程的求解方式,但由于其严重依赖基于反向传播的梯度下降和自动微分(AD),在计算效率和非凸优化方面面临重大瓶颈。
本文提出了“物理信息随机配置机”(PI-SCM),这是一个全新的、无需反向传播的框架,旨在解决微分方程的正向和反向问题。通过对非线性微分算子进行局部雅可比矩阵的解析评估,该框架将物理损失线性化并投影到一个统一的代数子空间中。这种方法使得最优网络权重可以通过广义线性最小二乘求解器显式确定,从而完全绕过了传统的非线性优化器。作者提出了一套渐进式算法(局部构建、滑动窗口更新和全局更新),并辅以通用逼近证明,展示了该方法在保持稳健预测精度的同时,实现了数量级上的训练加速。
📋 摘要
物理信息神经网络(PINNs)彻底改变了复杂微分方程的求解方式,但由于其严重依赖基于反向传播的梯度下降和自动微分(AD),在计算效率和非凸优化方面面临重大瓶颈。
Physics-Informed Neural Networks (PINNs) have transformed how complex differential equations are solved, but they suffer from major computational bottlenecks and non-convex optimization challenges due to their reliance on backpropagation-based gradient descent and automatic differentiation (AD).
为了解决这些局限性,本文引入了“物理信息随机配置机”(PI-SCM)——这是一个全新的、无需反向传播的框架,专为微分方程的正向和反向问题而设计。通过解析评估非线性微分算子的局部雅可比矩阵,该框架将物理损失线性化并将其投影到一个统一的代数子空间中。这使得最优网络权重可以通过广义线性最小二乘求解器显式确定,完全绕过了传统的非线性优化器。作者提出了一套渐进式算法(局部构建、滑动窗口更新和全局更新),并辅以通用逼近证明,证明了该方法在实现数量级更快训练的同时,具备稳健的预测精度。
To address these limitations, this paper introduces the Physics-Informed Stochastic Configuration Machine (PI-SCM)—a novel, backpropagation-free framework designed for both forward and inverse differential equation problems. By analytically evaluating local Jacobians for nonlinear differential operators, the framework linearizes the physical loss and projects it into a unified algebraic subspace. This allows optimal network weights to be determined explicitly via generalized linear least squares solvers, entirely bypassing traditional nonlinear optimizers. The authors propose a progressive algorithmic suite (localized construction, sliding-window updating, and global updating) backed by universal approximation proofs, demonstrating orders-of-magnitude faster training alongside robust predictive accuracy.
📚 文章元数据
- 主要学科: 数值分析 (
math.NA) - 次要学科: 人工智能 (
cs.AI), 机器学习 (cs.LG), 系统与控制 (eess.SY) - 备注: 17页,4幅图
- DOI: 10.48550/arXiv.2608.26549
- Primary Subject: Numerical Analysis (
math.NA)- Secondary Subjects: Artificial Intelligence (
cs.AI), Machine Learning (cs.LG), Systems and Control (eess.SY)- Comments: 17 pages, 4 figures
- DOI: 10.48550/arXiv.2608.26549