We propose a novel approximation hierarchy for cardinality-constrained, convex quadratic programs that exploits the rank-dominating eigenvectors of the quadratic matrix. Each level of approximation admits a min-max characterization whose objective function can be optimized over the binary variables analytically, while preserving convexity in the continuous variables. Exploiting this property, we propose two scalable optimization algorithms, coined as the “best response" and the “dual program", that can efficiently screen the potential indices of the nonzero elements of the original program. We show that the proposed methods are competitive with the existing screening methods in the current sparse regression literature, and it is particularly fast on instances with high number of measurements in experiments with both synthetic and real datasets.

Overparameterization refers to the important phenomenon where the width of a neural network is chosen such that learning algorithms can provably attain zero loss in nonconvex training. The existing theory establishes such global convergence using various initialization strategies, training modifications, and width scalings. In particular, the state-of-the-art results require the width to scale quadratically with the number of training data under standard initialization strategies used in practice for best generalization performance. In contrast, the most recent results obtain linear scaling either with requiring initializations that lead to the "lazy-training", or training only a single layer. In this work, we provide an analytical framework that allows us to adopt standard initialization strategies, possibly avoid lazy training, and train all layers simultaneously in basic shallow neural networks while attaining a desirable subquadratic scaling on the network width. We achieve the desiderata via Polyak-Lojasiewicz condition, smoothness, and standard assumptions on data, and use tools from random matrix theory.

Compressed sensing applied to magnetic resonance imaging (MRI) allows to reduce the scanning time by enabling images to be reconstructed from highly undersampled data. In this paper, we tackle the problem of designing a sampling mask for an arbitrary reconstruction method and a limited acquisition budget. Namely, we look for an optimal probability distribution from which a mask with a fixed cardinality is drawn. We demonstrate that this problem admits a compactly supported solution, which leads to a deterministic optimal sampling mask. We then propose a stochastic greedy algorithm that (i) provides an approximate solution to this problem, and (ii) resolves the scaling issues of [1], [2]. We validate its performance on in vivo dynamic MRI with retrospective undersampling, showing that our method preserves the performance of [1], [2] while reducing the computational burden by a factor close to 200. Our implementation is available at https://github.com/t-sanchez/stochasticGreedyMRI.

Linear networks provide valuable insights into the workings of neural networks in general. This paper identifies conditions under which the gradient flow provably trains a linear network, in spite of the non-strict saddle points present in the optimization landscape. This paper also provides the computational complexity of training linear networks with gradient flow. To achieve these results, this work develops a machinery to provably identify the stable set of gradient flow, which then enables us to improve over the state of the art in the literature of linear networks (Bah et al., 2019;Arora et al., 2018a). Crucially, our results appear to be the first to break away from the lazy training regime which has dominated the literature of neural networks. This work requires the network to have a layer with one neuron, which subsumes the networks with a scalar output, but extending the results of this theoretical work to all linear networks remains a challenging open problem.

A well-known first-order method for sampling from log-concave probability distributions is the Unadjusted Langevin Algorithm (ULA). This work proposes a new annealing step-size schedule for ULA, which allows to prove new convergence guarantees for sampling from a smooth log-concave distribution, which are not covered by existing state-of-the-art convergence guarantees. To establish this result, we derive a new theoretical bound that relates the Wasserstein distance to total variation distance between any two log-concave distributions that complements the reach of Talagrand $T_2$ inequality. Moreover, applying this new step size schedule to an existing constrained sampling algorithm, we show state-of-the-art convergence rates for sampling from a constrained log-concave distribution, as well as improved dimension dependence.

This paper introduces Memory-limited Online Subspace Estimation Scheme (MOSES) for both estimating the principal components of streaming data and reducing its dimension. More specifically, in various applications such as sensor networks, the data vectors are presented sequentially to a user who has limited storage and processing time available. Applied to such problems, MOSES can provide a running estimate of leading principal components of the data that has arrived so far and also reduce its dimension. MOSES generalises the popular incremental Singular Vale Decomposition (iSVD) to handle thin blocks of data, rather than just vectors. This minor generalisation in part allows us to complement MOSES with a comprehensive statistical analysis, thus providing the first theoretically-sound variant of iSVD, which has been lacking despite the empirical success of this method. This generalisation also enables us to concretely interpret MOSES as an approximate solver for the underlying non-convex optimisation program. We find that MOSES consistently surpasses the state of the art in our numerical experiments with both synthetic and real-world datasets, while being computationally inexpensive.

In this work, we propose a (linearized) Alternating Direction Method-of-Multipliers (ADMM) algorithm for minimizing a convex function subject to a nonconvex constraint. We focus on the special case where such constraint arises from the specification that a variable should lie in the range of a neural network. This is motivated by recent successful applications of Generative Adversarial Networks (GANs) in tasks like compressive sensing, denoising and robustness against adversarial examples. The derived rates for our algorithm are characterized in terms of certain geometric properties of the generator network, which we show hold for feedforward architectures, under mild assumptions. Unlike gradient descent (GD), it can efficiently handle non-smooth objectives as well as exploit efficient partial minimization procedures, thus being faster in many practical scenarios.

We propose a practical inexact augmented Lagrangian method (iALM) for nonconvex problems with nonlinear constraints. We characterize the total computational complexity of our method subject to a verifiable geometric condition, which is closely related to the Polyak-Lojasiewicz and Mangasarian-Fromowitz conditions. In particular, when a first-order solver is used for the inner iterates, we prove that iALM finds a first-order stationary point with $\tilde{\mathcal{O}}(1/\epsilon^3)$ calls to the first-order oracle. {If, in addition, the problem is smooth and} a second-order solver is used for the inner iterates, iALM finds a second-order stationary point with $\tilde{\mathcal{O}}(1/\epsilon^5)$ calls to the second-order oracle. These complexity results match the known theoretical results in the literature. We also provide strong numerical evidence on large-scale machine learning problems, including the Burer-Monteiro factorization of semidefinite programs, and a novel nonconvex relaxation of the standard basis pursuit template. For these examples, we also show how to verify our geometric condition.

Linear subspace models are pervasive in computational sciences and particularly used for large datasets which are often incomplete due to privacy issues or sampling constraints. Therefore, a critical problem is developing an efficient algorithm for detecting low-dimensional linear structure from incomplete data efficiently, in terms of both computational complexity and storage. In this paper we propose a streaming subspace estimation algorithm called Subspace Navigation via Interpolation from Partial Entries (SNIPE) that efficiently processes blocks of incomplete data to estimate the underlying subspace model. In every iteration, SNIPE finds the subspace that best fits the new data block but remains close to the previous estimate. We show that SNIPE is a streaming solver for the underlying nonconvex matrix completion problem, that it converges globally {to a stationary point of this program} regardless of initialization, and that the convergence is locally linear with high probability. We also find that SNIPE shows state-of-the-art performance in our numerical simulations.

Principal Component Analysis (PCA) finds the best linear representation for data and is an indispensable tool in many learning tasks. Classically, principal components of a dataset are interpreted as the directions that preserve most of its "energy", an interpretation that is theoretically underpinned by the celebrated Eckart-Young-Mirsky Theorem. There are yet other ways of interpreting PCA that are rarely exploited in practice, largely because it is not known how to reliably solve the corresponding non-convex optimisation programs. In this paper, we consider one such interpretation of principal components as the directions that preserve most of the "volume" of the dataset. Our main contribution is a theorem that shows that the corresponding non-convex program has no spurious local optima, and is therefore amenable to many convex solvers. We also confirm our findings numerically.

An incoherent low-rank matrix can be efficiently reconstructed after observing a few of its entries at random, and then, solving a convex program that minimizes the nuclear norm. In many applications, in addition to these entries, potentially valuable prior knowledge about the column and row spaces of the matrix is also available to the practitioner. In this paper, we incorporate this prior knowledge in matrix completion-by minimizing a weighted nuclear norm-and precisely quantify any improvements. In particular, we find in theory that reliable prior knowledge reduces the sample complexity of matrix completion by a logarithmic factor, and the observed improvement in numerical simulations is considerably more magnified. We also present similar results for the closely related problem of matrix recovery from generic linear measurements.

Consider an incoming sequence of vectors, all belonging to an unknown subspace S, and each with many missing entries. In order to estimate S, it is common to partition the data into blocks and iteratively update the estimate of S with each new incoming measurement block. In this letter, we investigate a rather basic question: Is it possible to identify S by averaging the range of the partially observed incoming measurement blocks on the Grassmannian? We show that, in general, the span of the incoming blocks is in fact a biased estimator of S when data suffer from erasures, and we find an upper bound for this bias. We reach this conclusion by examining the defining optimization program for the Frechet expectation on the Grassmannian, and with the aid of a sharp perturbation bound and standard large deviation results.

In this paper, we study a simple correlation-based strategy for estimating the unknown delay and amplitude of a signal based on a small number of noisy, randomly chosen frequency-domain samples. We model the output of this “compressive matched filter” as a random process whose mean equals the scaled, shifted autocorrelation function of the template signal. Using tools from the theory of empirical processes, we prove that the expected maximum deviation of this process from its mean decreases sharply as the number of measurements increases, and we also derive a probabilistic tail bound on the maximum deviation. Putting all of this together, we bound the minimum number of measurements required to guarantee that the empirical maximum of this random process occurs sufficiently close to the true peak of its mean function. We conclude that for broad classes of signals, this compressive matched filter will successfully estimate the unknown delay (with high probability and within a prescribed tolerance) using a number of random frequency-domain samples that scales inversely with the signal-to-noise ratio and only logarithmically in the observation bandwidth and the possible range of delays.

In the last few years, we have witnessed an explosion in applications of sparse representation, the majority of which share the need for finding sparse solutions of underdetermined systems of linear equations (USLE's). Based on recently proposed smoothed lscr0-norm (SL0), we develop a noise-tolerant algorithm for sparse representation, namely Robust-SL0, enjoying the same computational advantages of SL0, while demonstrating remarkable robustness against noise. The proposed algorithm is developed by adopting the corresponding optimization problem for noisy settings, followed by theoretically-justified approximation to reduce the complexity. Stability properties of robust-SL0 are rigorously analyzed, both analytically and experimentally, revealing a remarkable improvement in performance over SL0 and other competing algorithms, in the presence of noise.

The theory of compressive sampling involves making random linear projections of a signal. Provided signal is sparse in some basis, small number of such measurements preserves the information in the signal, with high probability. Following the success in signal reconstruction, compressive framework has recently proved useful in classification. In this paper, conventional random projection scheme is first extended to the image domain and the key notion of concentration of measure is studied. Findings are then employed to develop a 2D compressive classifier (2D-CC) for sparse images. Finally, theoretical results are validated within a realistic experimental framework.