430.758 (003) Topics in Signal Processing (Computational Imaging with Optimization and Deep Learning) Sample Midterm Problems Se Young Chun, PhD Associate Professor Department of Electrical & Computer Engineering Seoul National University April 12, 2022 Intelligent Computational imaging Lab (ICL) Q. Show the details: If f and g are Lipschitz continuous functions on R, then h(x) = f(x)g(x) is a Lipschitz continuous function on R. Q. For f(x) = (1/2)∥Ax − y∥22 we have ∥∇f(x) − ∇f(z)∥2 = ∥Aʹ(Ax − y) − Aʹ(Az − y)∥2 = ∥AʹA(x − z)∥2 ≤ ∥AʹA∥2 ∥x − z∥2. Then, find a Lipschitz constant of ∇f. Discuss whether your answer is the best Lipschitz constant for that ∇f or not. Q. The Fair potential is ψ(z) = δ2 (|z/δ| − log(1 + |z/δ|)) for some δ > 0. Is it Lipschitz continuous? Find the Lipschitz constant of the derivative of it. Q. Consider the complex LS cost function Ψ(x) = (1/2)∥Ax − y∥2 , where A ∈ CM×N and x ∈ CN and y ∈ CM . This function is not differentiable on CN. Show that g = Aʹ(Ax − y) is a best ascent direction for Ψ at x. (Hint: the equality condition for Cauchy-Schwarz inequality?) Q. Assume that the definition for the proximal operator is given. Determine analytically the proximal operator for the 2-norm. Then, determine the proximal operator for the (nonconvex) 0-norm. Q. Solve the LASSO problem as a constrained LS problem using the gradient projection (GP) method. (Hint: one can use the positive and negative parts of x separately). Intelligent Computational imaging Lab (ICL) 2 Q. Descriptions on the gradient descent, the steepest descent, the heavy ball method, the Nesterov’s fast gradient method, the Newton’s method (no more memorization than these. OGM is too much for memorizing, but I may ask something about OGM by providing related formular). Q. For the LS problem with Ψ(x) = (1/2)∥Ax − y∥2, describe the ideal preconditioner using ordinary GD and preconditioned GD. Why is it the ideal preconditioner? Q. See Lecture 6, page 11 (a quadratic majorizer for the LASSO problem). For the quadratic majorizer for (1/2)∥Ax − y∥2, find c1 and c2. Q. Describe the ISTA, the proximal gradient method (PGM) and the IHT. Q. For the cost function of the form f(x) = ∑$ !"# hi([Ax]i), derive a quadratic majorizer of it using additive separability assuming that hi are convex. Q. Derive the ML-EM using MM approach. Q. Describe the differences between maximum curvature and optimal curvature (Huber’s majorizer). Intelligent Computational imaging Lab (ICL) 3 Q. Consider f(x) = 3 ∥x∥1+XC(x), C={x ∈ RN: ∥x∥∞ ≤5}. Find the proximal operator of f. Describe the proximal point algorithm. Q. Describe the fast proximal gradient method (FPGM) or FISTA (POGM is too complicated to memorize). Q. Find the block coordinate minimization (BCM) for the following joint cost function: (1/2)∥Ax − y∥22 + b ((1/2)∥x − D z∥22 + c ∥z∥1). Find the BCD for the same problem. Q. Given N × L data matrix X, consider the following dictionary learning optimization problem: argminD minZ Ψ(D,Z), Ψ(D,Z)=(1/2)|X−DZ|F2+β∥vec(Z)∥0, where D is the set of N × K matrices having unit-norm columns. The drawback of the atom-wise BCM approach is that it is not conducive to parallel computing. (a) Write down a MM (or equivalent PGM) update for D. (This update is mentioned in the course notes without any formula given.) (b) An alternative is to make a BCD approach where one updates, say, two atoms of D together. Thinking of the partition D = [D1:2 D3:K ] write a concise expression for an update of D1:2 that is guaranteed to decrease Ψ. Discuss what advantage this approach has over (a). Intelligent Computational imaging Lab (ICL) 4 Q. Describe the variable splitting, the Lagrangian function, and augmented Lagrangian method. Q. Find the convex conjugate of the affine function. Q. Examples in Lecture notes 12, 13. Q. Describe the Lagrangian dual function, the duality gap and the ADMM. Q. Describe the linearized augmented Lagrangian method. Q. Describe the primal-dual hybrid gradient method. Intelligent Computational imaging Lab (ICL) 5 Questions Any questions? sychun@snu.ac.kr Intelligent Computational imaging Lab (ICL) 6