Presentation is loading. Please wait.

Presentation is loading. Please wait.

NESTA: A Fast and Accurate First-Order Method for Sparse Recovery

Similar presentations


Presentation on theme: "NESTA: A Fast and Accurate First-Order Method for Sparse Recovery"— Presentation transcript:

1 NESTA: A Fast and Accurate First-Order Method for Sparse Recovery
Reading Group NESTA: A Fast and Accurate First-Order Method for Sparse Recovery (STEPHEN BECKER, JEROME BOBIN and EMMANUEL J. CANDES) Presenter: Feng Lin Cognitive Radio Institute ECE / CMR Tennessee Technological University March 18, 2011

2 Outline Introduction Nesterov’s method NESTA algorithm details
Accelerate NESTA with continuation Accurate optimization Comparison among other algorithms Problem solving applications Conclusion 11/17/2018

3 Motivation Solving large-scale problems is challenging
Standard second-order methods are accurate but problematic to compute lots of newton steps. A lot of first order methods may be faster but not accurate. So we would like to pursue the high accuracy with more digits of precision Another motivation is solving signal is not exactly sparse, but rather approximately sparse, as in real world compressible signals. 11/17/2018

4 Advantage of NESTA Ideally suited for solving large-scale compressed sensing reconstruction problems 1) It is computationally efficient 2) It is accurate and returns solutions with several correct digits 3) It is flexible and amenable to many kinds of reconstruction problems 4) It is robust in the sense that its excellent performance across a wide range of problems does not depend on the fine tuning of several parameters 11/17/2018

5 Introduction Signal model of Compressed sensing 11/17/2018

6 Introduction Some optimization problems that will be discussed throughout Quadratically constrained l1-minimization problem Equivalent Lagrangian form Lasso 11/17/2018

7 Nesterov’s method Minimizing smooth convex functions
Function f is assumed to be differentiable and its gradient is Lipschitz and obeys L is an upper bound on Lipschitz constant Nesterov's algorithm minimizes f(x) over Qp(primal feasible set) by iteratively estimating three sequences xk, yk and zk while smoothing the feasible set Qp. 11/17/2018

8 Nesterov’s method Minimizing nonsmooth convex functions. Recast as
Nesterov proposed smooth method is Lipschitz with constant 11/17/2018

9 Yk is the current guess of the optimal solution.
Zk involves a weighted sum of already computed gradients from theoretical analysis of other papers. 11/17/2018

10 Nesterov’s method : Convergence rate 11/17/2018

11 NESTA The algorithm uses two ideas due to Yurii Nesterov.
The first idea is an accelerated convergence scheme for first-order methods, giving the optimal convergence rate for this class of problems. The second idea is a smoothing technique that replaces the non-smooth l1 norm with a smooth version. We assume that A*A is an orthogonal project, i.e. the rows of A are orthogonal. 11/17/2018

12 NESTA Target: we wish to solve the BP problem Form smooth function:
11/17/2018

13 Lagrangian for this problem is
Updating yk: Where Lagrangian for this problem is We can get the KKT conditions 11/17/2018

14 Solve the KKT function, we get
Updating zk We may choose As before, we can get 11/17/2018

15 Parameters selection A single smoothing parameter µ
A suitable stopping criterion Terminate the algorithm when the relative variation of fµ is small, i.e , where , depending upon the desired accuracy. 11/17/2018

16 Accelerate NESTA with continuation
Continuation is a very useful tool to increase the speed of convergence, in particular when dealing with large-scale problems and high dynamic range signals. For NESTA algorithm, the convergence rate obeys Since Lµ is proportional to 1/µ, the larger µ and the faster the convergence. Choosing a good guess x0 close to xµ also improving the rate of convergence, because low value of 11/17/2018

17 Interpretation 11/17/2018

18 11/17/2018

19 Continuation step T = 4 , 5 ,6 leads to reasonable results.
11/17/2018

20 Accurate optimization
Two criteria Relative error on the objective function Accuracy of the optimal solution itself FISTA NESTA comparison 11/17/2018

21 µ and accuracy When µ decreases,
the accuracy of NESTA increases, decreasing µ by a factor of 10 gives about 1 additional digit of accuracy on the optimal value. NA increase accordingly. µ = 0.02 seems a reasonable choice to balance these two considerations 11/17/2018

22 Comparison among other algorithms
NESTA Gradient Projections for Sparse Reconstruction (GPSR) Sparse reconstruction by separable approximation (SpaRSA) l1 regularized least squares (l1_ls) Spectral projected gradient (SPGL1) Fixed Point Continuation method (FPC) FPC Active Set (FPC-AS) Bregman Fast Iterative Soft-Thresholding Algorithm (FISTA) 11/17/2018

23 Constrained versus unconstrained minimization
NESTA solve the constrained problem (BP) SPGL1 solve the constrained problem (LST) while all other methods tested solve the unconstrained problem (QP). In general, solving an unconstrained problem is much easier than a constrained problem. 11/17/2018

24 11/17/2018

25 11/17/2018

26 Application : Nonstandard sparse reconstruction: l1 analysis
For real-world signals, the signal is not sparse, but it may be sparse (or be approximately-sparse, i.e. most energy is concentrated in only a few coefficients) in a dictionary W, can be represented as X=Wα. Traditionally, we attempts reconstruction by solving synthesis-based approach, simply treat (AW) as a single operator. If W is orthonormal, it is equivalent to We can apply NESTA, only need to change the step1, while step 2, 3 remains 11/17/2018

27 Application : Total-variation minimization
The TV norm of a 2D digital object x[i; j] is given where D1 and D2 are the horizontal and vertical differences We also can form The rest are the same 11/17/2018

28 Extension This paper focused on the situation in which A*A is a projector (the rows of A are orthonormal). Most computationally friendly compressed sensing are of this form It allows fast computations of the two sequence of iterates yk, zk. However, NESTA can cope with the problem with A*A is not a projection (or not diagonal). We need to solve Then A good rule for selecting reasonable λ is critical for extending NESTA to a general problem 11/17/2018

29 Conclusions NESTA , an iterative algorithm with fast convergence rate based on Nesterov’s algorithm, is a method of choice for solving large-scale problems. NESTA is accurate that can find the first 4 or 5 significant digits of the optimal solution. NESTA can be adapted to solve many problems beyond l1 minimization with the same efficiency, such as l1 analysis, total-variation (TV) minimization problems. 11/17/2018

30 Thank you! 11/17/2018


Download ppt "NESTA: A Fast and Accurate First-Order Method for Sparse Recovery"

Similar presentations


Ads by Google