卷积神经网络的最优推断控制
Optimal Inferential Control of Convolutional Neural Networks
AI总结:
针对CNNs超高维数和强非线性导致传统梯度最优控制失效的挑战,提出最优推断控制框架,并开发基于矩阵变量分布的序列集合卡尔曼平滑器以利用GPU加速计算,有效实现了对高维时空系统的控制。
摘要:
卷积神经网络(CNNs)在蓬勃发展的科学机器学习领域中,在表示和模拟复杂的时空动态系统方面取得了显著成功。然而,CNNs的最优控制提出了严峻挑战,因为CNNs固有的超高维数和强非线性使其难以应用传统的基于梯度的最优控制技术。为应对这一挑战,我们提出了一种针对表示复杂时空系统的CNNs的最优推断控制框架,该框架根据指定的控制目标顺序推断最佳控制决策。这种重构使得序列蒙特卡洛采样得以应用,该方法在非线性推断的高维空间搜索中十分高效。我们具体利用集合卡尔曼平滑(一种序列蒙特卡洛算法),以发挥其在非线性高维系统中的计算效率。此外,为了利用图形处理器(GPUs)加速计算,我们开发了一种基于矩阵变量分布的新型序列集合卡尔曼平滑器。该平滑器能够直接处理CNNs基于矩阵的输入和输出,无需向量化即可适应GPU的并行计算架构。数值实验表明,所提方法在控制具有高维状态和控制空间的时空系统方面是有效的。所有代码和数据可在 https://github.com/Alivaziri/Optimal-Inferential-Control-of-CNNs 获取。
英文摘要:
Convolutional neural networks (CNNs) have achieved remarkable success in representing and simulating complex spatio-temporal dynamic systems within the burgeoning field of scientific machine learning. However, optimal control of CNNs poses a formidable challenge, because the ultra-high dimensionality and strong nonlinearity inherent in CNNs render them resistant to traditional gradient-based optimal control techniques. To tackle the challenge, we propose an optimal inferential control framework for CNNs that represent a complex spatio-temporal system, which sequentially infers the best control decisions based on the specified control objectives. This reformulation opens up the utilization of sequential Monte Carlo sampling, which is efficient in searching through high-dimensional spaces for nonlinear inference. We specifically leverage ensemble Kalman smoothing, a sequential Monte Carlo algorithm, to take advantage of its computational efficiency for nonlinear high-dimensional systems. Further, to harness graphics processing units (GPUs) to accelerate the computation, we develop a new sequential ensemble Kalman smoother based on matrix variate distributions. The smoother is capable of directly handling matrix-based inputs and outputs of CNNs without vectorization to fit with the parallelized computing architecture of GPUs. Numerical experiments show that the proposed approach is effective in controlling spatio-temporal systems with high-dimensional state and control spaces. All the code and data are available at https://github.com/Alivaziri/Optimal-Inferential-Control-of-CNNs.
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。