Stochastic primal-dual three operator splitting algorithm with extension to equivariant regularization-by-denoising
Applications of operator theory in optimization, convex analysis, mathematical programming, economics (47N10) Numerical methods involving duality (49M29) Numerical solutions of ill-posed problems in abstract spaces; regularization (65J20) Numerical optimization and variational techniques (65K10) Computing methodologies for image processing (68U10) Stochastic programming (90C15) Image processing (compression, reconstruction, etc.) in information and communication theory (94A08)
This interesting paper discusses a new stochastic primal-dual three operator splitting algorithm with extensions to equivariant regularization via denoising. To set the scene, the authors introduce the following algorithm: \N\[\Nx^*=\min_{x\in X}(f(Ax)+g(x)+h(x)), \N\] \Nwhere \(X\in \mathbb R^d\) is convex, \(A\in\mathbb R^{N\times d}\) is a linear operator, \(f(A\cdot)\) represents data fidelity, \(g\) is a regularizer which admits a sample proximal operator, and \(h\) is another regularizer for which gradients are readily acessible. All three terms, \(f\), \(g\), \(h\), are proper convex lower-semicontinuous functions, while in addition \(h\) has Lipschitz-continuous gradients. The problem is often formulated as a dual saddle-point optimizer problem of the form:\N\[\N[x^*,y^*]=\min_{x\in X}\max_{y\in Y}(h(x)+g(x)+\langle Ax,y \rangle -f^*(y)),\N\]\Nwhere \(f^*(y)=\sup_{z\in Y}\langle z,y \rangle -f(z)\). In this paper the authors use stochastic gradient methods to solve this latter saddle-point optimization problem if it admits a finite-sum structure. Their algorithm in this regard shows an ergodic \(O(1/K)\) convergence rate. The authors also provide extensions of their algorithm, which have applications in several imaging, denoising and inverse problems.\N\NFor the entire collection see [Zbl 1573.68017].
- A generalized forward-backward splitting
- A primal-dual splitting method for convex optimization involving Lipschitzian, proximable and linear composite terms
- A proximal stochastic gradient method with progressive variance reduction
- A splitting algorithm for dual monotone inclusions involving cocoercive operators
- A three-operator splitting scheme and its optimization applications
- Accelerated, parallel, and proximal coordinate descent
- Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising
- Iteration complexity of randomized block-coordinate descent methods for minimizing a composite function
- Katyusha: the first direct acceleration of stochastic gradient methods
- On the ergodic convergence rates of a first-order primal-dual algorithm
- Stochastic primal-dual hybrid gradient algorithm with adaptive step sizes
- Stochastic Primal-Dual Hybrid Gradient Algorithm with Arbitrary Sampling and Imaging Applications
- The Little Engine that Could: Regularization by Denoising (RED)
This page was built for publication: Stochastic primal-dual three operator splitting algorithm with extension to equivariant regularization-by-denoising
Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q6908298)
