It has been shown that the problem of \ell_1 -penalized least-square regression commonly referred to as the Lasso or Basis Pursuit DeNoising leads to solutions that are sparse and therefore achieves model selection. We propose in this paper an algorithm to solve the Lasso with online observations. We introduce an optimization problem that allows us to compute an homotopy from the current solution to the solution after observing a new data point. We compare our method to Lars and present an application to compressed sensing with sequential observations. Our approach can also be easily extended to compute an homotopy from the current solution to the solution after removing a data point, which leads to an efficient algorithm for leave-one-out cross-validation.

The problems of Lasso regression and optimal design of experiments share a critical property: their optimal solutions are typically \emph{sparse}, i.e., only a small fraction of the optimal variables are non-zero. Therefore, the identification of the support of an optimal solution reduces the dimensionality of the problem and can yield a substantial simplification of the calculations. It has recently been shown that linear regression with a \emph{squared} $\ell_1$-norm sparsity-inducing penalty is equivalent to an optimal experimental design problem. In this work, we use this equivalence to derive safe screening rules that can be used to discard inessential samples. Compared to previously existing rules, the new tests are much faster to compute, especially for problems involving a parameter space of high dimension, and can be used dynamically within any iterative solver, with negligible computational overhead. Moreover, we show how an existing homotopy algorithm to compute the regularization path of the lasso method can be reparametrized with respect to the squared $\ell_1$-penalty. This allows the computation of a Bayes $c$-optimal design in a finite number of steps and can be several orders of magnitude faster than standard first-order algorithms. The efficiency of the new screening rules and of the homotopy algorithm are demonstrated on different examples based on real data.

This paper is concerned with the statistical analysis of matrix-valued time series. These are data collected over a network of sensors (typically a set of spatial locations), recording, over time, observations of multiple measurements. From such data, we propose to learn, in an online fashion, a graph that captures two aspects of dependency: one describing the sparse spatial relationship between sensors, and the other characterizing the measurement relationship. To this purpose, we introduce a novel multivariate autoregressive model to infer the graph topology encoded in the coefficient matrix which captures the sparse Granger causality dependency structure present in such matrix-valued time series. We decompose the graph by imposing a Kronecker sum structure on the coefficient matrix. We develop two online approaches to learn the graph in a recursive way. The first one uses Wald test for the projected OLS estimation, where we derive the asymptotic distribution for the estimator. For the second one, we formalize a Lasso-type optimization problem. We rely on homotopy algorithms to derive updating rules for estimating the coefficient matrix. Furthermore, we provide an adaptive tuning procedure for the regularization parameter. Numerical experiments using both synthetic and real data, are performed to support the effectiveness of the proposed learning approaches.

Sparse adversarial attacks can fool deep neural networks (DNNs) by only perturbing a few pixels (regularized by l_0 norm). Recent efforts combine it with another l_infty imperceptible on the perturbation magnitudes. The resultant sparse and imperceptible attacks are practically relevant, and indicate an even higher vulnerability of DNNs that we usually imagined. However, such attacks are more challenging to generate due to the optimization difficulty by coupling the l_0 regularizer and box constraints with a non-convex objective. In this paper, we address this challenge by proposing a homotopy algorithm, to jointly tackle the sparsity and the perturbation bound in one unified framework. Each iteration, the main step of our algorithm is to optimize an l_0-regularized adversarial loss, by leveraging the nonmonotone Accelerated Proximal Gradient Method (nmAPG) for nonconvex programming; it is followed by an l_0 change control step, and an optional post-attack step designed to escape bad local minima. We also extend the algorithm to handling the structural sparsity regularizer. We extensively examine the effectiveness of our proposed homotopy attack for both targeted and non-targeted attack scenarios, on CIFAR-10 and ImageNet datasets. Compared to state-of-the-art methods, our homotopy attack leads to significantly fewer perturbations, e.g., reducing 42.91% on CIFAR-10 and 75.03% on ImageNet (average case, targeted attack), at similar maximal perturbation magnitudes, when still achieving 100% attack success rates. Our codes are available at: https://github.com/VITA-Group/SparseADV_Homotopy.

Nonnegative least squares (NNLS) problems arise in models that rely on additive linear combinations. In particular, they are at the core of nonnegative matrix factorization (NMF) algorithms. The nonnegativity constraint is known to naturally favor sparsity, that is, solutions with few non-zero entries. However, it is often useful to further enhance this sparsity, as it improves the interpretability of the results and helps reducing noise. While the $\ell_0$-"norm", equal to the number of non-zeros entries in a vector, is a natural sparsity measure, its combinatorial nature makes it difficult to use in practical optimization schemes. Most existing approaches thus rely either on its convex surrogate, the $\ell_1$-norm, or on heuristics such as greedy algorithms. In the case of multiple right-hand sides NNLS (MNNLS), which are used within NMF algorithms, sparsity is often enforced column- or row-wise, and the fact that the solution is a matrix is not exploited. In this paper, we first introduce a novel formulation for sparse MNNLS, with a matrix-wise $\ell_0$ sparsity constraint. Then, we present a two-step algorithm to tackle this problem. The first step uses a homotopy algorithm to produce the whole regularization path for all the $\ell_1$-penalized NNLS problems arising in MNNLS, that is, to produce a set of solutions representing different tradeoffs between reconstruction error and sparsity. The second step selects solutions among these paths in order to build a sparsity-constrained matrix that minimizes the reconstruction error. We illustrate the advantages of our proposed algorithm for the unmixing of facial and hyperspectral images.

Most of the existing methods for sparse signal recovery assume a static system: the unknown signal is a finite-length vector for which a fixed set of linear measurements and a sparse representation basis are available and an L1-norm minimization program is solved for the reconstruction. However, the same representation and reconstruction framework is not readily applicable in a streaming system: the unknown signal changes over time, and it is measured and reconstructed sequentially over small time intervals. In this paper, we discuss two such streaming systems and a homotopy-based algorithm for quickly solving the associated L1-norm minimization programs: 1) Recovery of a smooth, time-varying signal for which, instead of using block transforms, we use lapped orthogonal transforms for sparse representation. 2) Recovery of a sparse, time-varying signal that follows a linear dynamic model. For both the systems, we iteratively process measurements over a sliding interval and estimate sparse coefficients by solving a weighted L1-norm minimization program. Instead of solving a new L1 program from scratch at every iteration, we use an available signal estimate as a starting point in a homotopy formulation. Starting with a warm-start vector, our homotopy algorithm updates the solution in a small number of computationally inexpensive steps as the system changes. The homotopy algorithm presented in this paper is highly versatile as it can update the solution for the L1 problem in a number of dynamical settings. We demonstrate with numerical experiments that our proposed streaming recovery framework outperforms the methods that represent and reconstruct a signal as independent, disjoint blocks, in terms of quality of reconstruction, and that our proposed homotopy-based updating scheme outperforms current state-of-the-art solvers in terms of the computation time and complexity.

To recover a sparse signal from an underdetermined system, we often solve a constrained L1-norm minimization problem. In many cases, the signal sparsity and the recovery performance can be further improved by replacing the L1 norm with a "weighted" L1 norm. Without any prior information about nonzero elements of the signal, the procedure for selecting weights is iterative in nature. Common approaches update the weights at every iteration using the solution of a weighted L1 problem from the previous iteration. In this paper, we present two homotopy-based algorithms that efficiently solve reweighted L1 problems. First, we present an algorithm that quickly updates the solution of a weighted L1 problem as the weights change. Since the solution changes only slightly with small changes in the weights, we develop a homotopy algorithm that replaces the old weights with the new ones in a small number of computationally inexpensive steps. Second, we propose an algorithm that solves a weighted L1 problem by adaptively selecting the weights while estimating the signal. This algorithm integrates the reweighting into every step along the homotopy path by changing the weights according to the changes in the solution and its support, allowing us to achieve a high quality signal reconstruction by solving a single homotopy problem. We compare the performance of both algorithms, in terms of reconstruction accuracy and computational complexity, against state-of-the-art solvers and show that our methods have smaller computational cost. In addition, we will show that the adaptive selection of the weights inside the homotopy often yields reconstructions of higher quality.

It has been shown that the problem of $\ell_1$-penalized least-square regression commonly referred to as the Lasso or Basis Pursuit DeNoising leads to solutions that are sparse and therefore achieves model selection. We propose in this paper an algorithm to solve the Lasso with online observations. We introduce an optimization problem that allows us to compute an homotopy from the current solution to the solution after observing a new data point. We compare our method to Lars and present an application to compressed sensing with sequential observations. Our approach can also be easily extended to compute an homotopy from the current solution to the solution after removing a data point, which leads to an efficient algorithm for leave-one-out cross-validation.