Deep learning as optimal control problems: models and numerical methods
From MaRDI portal
Publication:2297872
Existence theories for optimal control problems involving ordinary differential equations (49J15) Discrete approximations in optimal control (49M25) Numerical optimization and variational techniques (65K10) Numerical solution of boundary value problems involving ordinary differential equations (65L10) Numerical methods for Hamiltonian systems including symplectic integrators (65P10) Artificial neural networks and deep learning (68T07)
Abstract: We consider recent work of Haber and Ruthotto 2017 and Chang et al. 2018, where deep learning neural networks have been interpreted as discretisations of an optimal control problem subject to an ordinary differential equation constraint. We review the first order conditions for optimality, and the conditions ensuring optimality after discretisation. This leads to a class of algorithms for solving the discrete optimal control problem which guarantee that the corresponding discrete necessary conditions for optimality are fulfilled. The differential equation setting lends itself to learning additional parameters such as the time discretisation. We explore this extension alongside natural constraints (e.g. time steps lie in a simplex). We compare these deep learning algorithms numerically in terms of induced flow and generalisation ability.
Recommendations
- A mean-field optimal control formulation of deep learning
- Dynamical Systems andOptimal Control Approach to Deep Learning
- Optimal control by deep learning techniques and its applications on epidemic models
- Deep neural networks algorithms for stochastic control problems on finite horizon: numerical applications
- Selection dynamics for deep neural networks
Cites work
- A mean-field optimal control formulation of deep learning
- A proposal on machine learning via dynamical systems
- A Technique for the Numerical Solution of Certain Integral Equations of the First Kind
- Artificial intelligence as structural estimation: Deep Blue, Bonanza, and AlphaGo
- Clebsch optimal control formulation in mechanics
- Control theory from the geometric viewpoint.
- Deep learning: an introduction for applied mathematicians
- Deep limits of residual neural networks
- Geometric Numerical Integration
- scientific article; zbMATH DE number 1182386 (Why is no real title available?)
- scientific article; zbMATH DE number 1796939 (Why is no real title available?)
- scientific article; zbMATH DE number 845714 (Why is no real title available?)
- scientific article; zbMATH DE number 3227378 (Why is no real title available?)
- Linear Inversion of Band-Limited Reflection Seismograms
- Machine learning: deepest learning as statistical data assimilation problems
- Maximum principle based algorithms for deep learning
- Modern regularization methods for inverse problems
- Neural network with unbounded activation functions is universal approximator
- Nonlinear inverse scale space methods
- On the necessity of negative coefficients for operator splitting schemes of order higher than two
- Pattern recognition and machine learning.
- Runge-Kutta methods in optimal control and the transformed adjoint system
- Stable architectures for deep neural networks
- Symplectic Runge-Kutta schemes for adjoint equations, automatic differentiation, optimal control, and more
- The Euler approximation in state constrained optimal control
- Two-Point Boundary Value Problems of Linear Hamiltonian Systems
- Variational, Geometric, and Level Set Methods in Computer Vision
Cited in
(50)- Deep neural networks and mixed integer linear optimization
- A projected primal-dual gradient optimal control method for deep reinforcement learning
- Variational networks: an optimal control approach to early stopping variational methods for image restoration
- Classification with Runge-Kutta networks and feature space augmentation
- A measure theoretical approach to the mean-field maximum principle for training NeurODEs
- Deep learning for inverse problems. Abstracts from the workshop held March 7--13, 2021 (hybrid meeting)
- Preface. Special issue in honor of Reinout Quispel
- Deep relaxation: partial differential equations for optimizing deep neural networks
- A mean-field optimal control formulation of deep learning
- Linear-quadratic stochastic delayed control and deep learning resolution
- A framework for randomized time-splitting in linear-quadratic optimal control
- Control on the manifolds of mappings with a view to the deep learning
- Neural control of discrete weak formulations: Galerkin, least squares \& minimal-residual methods with quasi-optimal weights
- Optimal control by deep learning techniques and its applications on epidemic models
- An approach to solving optimal control problems of nonlinear systems by introducing detail-reward mechanism in deep reinforcement learning
- Maximum principle based algorithms for deep learning
- Understanding recurrent neural networks using nonequilibrium response theory
- Structure-preserving deep learning
- Algorithms for solving high dimensional PDEs: from nonlinear Monte Carlo to machine learning
- Deep neural networks, generic universal interpolation, and controlled ODEs
- Deep learning for ranking response surfaces with applications to optimal stopping problems
- Machine learning: deepest learning as statistical data assimilation problems
- Neural Approximations for Optimal Control and Decision
- Dynamical Systems andOptimal Control Approach to Deep Learning
- Turnpike in optimal control of PDEs, ResNets, and beyond
- Deep neural networks on diffeomorphism groups for optimal shape reparametrization
- Deep learning approximation of diffeomorphisms via linear-control systems
- Sparsity in long-time control of neural ODEs
- Accuracy Estimates for Bilinear Optimal Control Problems Governed by Ordinary Differential Equations
- Data-driven robust optimization using deep neural networks
- Neural ODE Control for Classification, Approximation, and Transport
- Geometric methods for adjoint systems
- An optimal time variable learning framework for deep neural networks
- Optimal Dirichlet boundary control by Fourier neural operators applied to nonlinear optics
- Efficient and stable SAV-based methods for gradient flows arising from deep learning
- PottsMGNet: a mathematical explanation of encoder-decoder based neural networks
- An optimal control framework for adaptive neural ODEs
- A descent algorithm for the optimal control of ReLU neural network informed PDEs based on approximate directional derivatives
- On properties of adjoint systems for evolutionary PDEs
- NINNs: Nudging induced neural networks
- Predict globally, correct locally: parallel-in-time optimization of neural networks
- Double-well net for image segmentation
- Constrained dynamics, stochastic numerical methods and the modeling of complex systems. Abstracts from the workshop held May 26--31, 2024
- Variational principles for Hamiltonian systems
- Random sampling-based gradient descent method for optimal control problems with variance reduction
- A mathematical explanation of UNet
- A type II Hamiltonian variational principle and adjoint systems for Lie groups
- From NeurODEs to AutoencODEs: a mean-field control framework for width-varying neural networks
- An optimal control approach for neural network architecture adaptation with a posteriori error estimation
- Machine learning from a continuous viewpoint. I
This page was built for publication: Deep learning as optimal control problems: models and numerical methods
Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q2297872)
