Title:Optimal Inferential Control of Convolutional Neural Networks

Abstract: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 this https URL.