Abstract:Sparse optimization involving the L0-norm function as the regularization in objective function has a wide application in many fields. In this paper, we propose a projected neural network modeled by a differential equation to solve a class of these optimization problems, in which the objective function is the sum of a nonsmooth convex loss function and the regularization defined by the L0-norm function. This optimization problem is not only nonconvex, but also discontinuous. To simplify the structure of the proposed network and let it own better convergence properties, we use the smoothing method, where the new constructed smoothing function for the regularization term plays a key role. We prove that the solution to the proposed network is globally existent and unique, and any accumulation point of it is a critical point of the continuous relaxation model. Except for a special case, which can be easily justified, any critical point is a local minimizer of the considered sparse optimization problem. It is an interesting thing that all critical points own a promising lower bound property, which is satisfied by all global minimizers of the considered problem, but is not by all local minimizers. Finally, we use some numerical experiments to illustrate the efficiency and good performance of the proposed method for solving this class of sparse optimization problems, which include the most widely used models in feature selection of classification learning.

L(1)-Regression Based Subdivision Schemes for Noisy Data

Convergence of Projected Subgradient Method with Sparse or Low-Rank Constraints

Progressive and Iterative Approximation for Least Squares B-Spline Curve and Surface Fitting

Subdivision Schemes Based on Weighted Local Polynomial Regression: A New Technique for the Convergence Analysis

Sparse Estimation Via ℓ_q Optimization Method in High-Dimensional Linear Regression

Optimal linear subdivision rules for noisy data

Discontinuity-Preserving Moving Least Squares Method

Subdivision schemes based on weighted local polynomial regression. A new technique for the convergence analysis

Families of non-linear subdivision schemes for scattered data fitting and their non-tensor product extensions

Robust Sparse Recovery in Impulsive Noise via lp-l1 Optimization

Developable surface reconstruction from noisy data with L0-norm minimization

A Robust Recovery Algorithm with Smoothing Strategies.

Low-Rank-Sparse Subspace Representation for Robust Regression

L-1/2 NORM REGULARIZATION FOR 3D SEISMIC DATA INTERPOLATION

l1-norm regularized l1-norm best-fit lines

A Data-Driven Line Search Rule for Support Recovery in High-Dimensional Data Analysis.

Iteratively Reweighted Sparse Reconstruction in Impulsive Noise

Robust Sparse Recovery Via Weakly Convex Optimization in Impulsive Noise

Neural Network for a Class of Sparse Optimization with L0-regularization.

Robust Sparse Recovery in Impulsive Noise Via $\ell _p$ -$\Ell _1$ Optimization

A new penalized least absolute deviation model for high dimensional sparse linear regression and an efficient sequential linear programming algorithm