COCO is a platform for Comparing Continuous Optimizers in a black-box setting. It aims at automatizing the tedious and repetitive task of benchmarking numerical optimization algorithms to the greatest possible extent. We present the rationals behind the development of the platform as a general proposition for a guideline towards better benchmarking. We detail underlying fundamental concepts of COCO such as its definition of a problem, the idea of instances, the relevance of target values, and runtime as central performance measure. Finally, we give a quick overview of the basic code structure and the available test suites.

In this paper, we study the proximal gradient algorithm with extrapolation for minimizing the sum of a Lipschitz differentiable function and a proper closed convex function. Under the error bound condition used in [19] for analyzing the convergence of the proximal gradient algorithm, we show that there exists a threshold such that if the extrapolation coefficients are chosen below this threshold, then the sequence generated converges $R$-linearly to a stationary point of the problem. Moreover, the corresponding sequence of objective values is also $R$-linearly convergent. In addition, the threshold reduces to $1$ for convex problems and, as a consequence, we obtain the $R$-linear convergence of the sequence generated by FISTA with fixed restart. Finally, we present some numerical experiments to illustrate our results.

When faced with a supervised learning problem, we hope to have rich enough data to build a model that predicts future instances well. However, in practice, problems can exhibit predictive heterogeneity: most instances might be relatively easy to predict, while others might be predictive outliers for which a model trained on the entire dataset does not perform well. Identifying these can help focus future data collection. We present gLOP, the global and Local Penalty, a framework for capturing predictive heterogeneity and identifying predictive outliers. gLOP is based on penalized regression for multitask learning, which improves learning by leveraging training signal information from related tasks. We give two optimization algorithms for gLOP, one space-efficient, and another giving the full regularization path. We also characterize uniqueness in terms of the data and tuning parameters, and present empirical results on synthetic data and on two health research problems.

We study Frank-Wolfe methods for nonconvex stochastic and finite-sum optimization problems. Frank-Wolfe methods (in the convex case) have gained tremendous recent interest in machine learning and optimization communities due to their projection-free property and their ability to exploit structured constraints. However, our understanding of these algorithms in the nonconvex setting is fairly limited. In this paper, we propose nonconvex stochastic Frank-Wolfe methods and analyze their convergence properties. For objective functions that decompose into a finite-sum, we leverage ideas from variance reduction techniques for convex optimization to obtain new variance reduced nonconvex Frank-Wolfe methods that have provably faster convergence than the classical Frank-Wolfe method. Finally, we show that the faster convergence rates of our variance reduced methods also translate into improved convergence rates for the stochastic setting.

Source imaging based on magnetoencephalography (MEG) and electroencephalography (EEG) allows for the non-invasive analysis of brain activity with high temporal and good spatial resolution. As the bioelectromagnetic inverse problem is ill-posed, constraints are required. For the analysis of evoked brain activity, spatial sparsity of the neuronal activation is a common assumption. It is often taken into account using convex constraints based on the l1-norm. The resulting source estimates are however biased in amplitude and often suboptimal in terms of source selection due to high correlations in the forward model. In this work, we demonstrate that an inverse solver based on a block-separable penalty with a Frobenius norm per block and a l0.5-quasinorm over blocks addresses both of these issues. For solving the resulting non-convex optimization problem, we propose the iterative reweighted Mixed Norm Estimate (irMxNE), an optimization scheme based on iterative reweighted convex surrogate optimization problems, which are solved efficiently using a block coordinate descent scheme and an active set strategy. We compare the proposed sparse imaging method to the dSPM and the RAP-MUSIC approach based on two MEG data sets. We provide empirical evidence based on simulations and analysis of MEG data that the proposed method improves on the standard Mixed Norm Estimate (MxNE) in terms of amplitude bias, support recovery, and stability.

School of Engineering, Brown University Providence, RI 02912, USA Abstract Non-convex optimization is ubiquitous in machine learning. The bound at each iteration is required to touch the objective function at the optimizer of the previous bound. We show that this touching constraint is unnecessary and overly restrictive. We generalize MM by relaxing this constraint, and propose a new optimization framework, named Generalized Majorization-Minimization (G-MM) that is more flexible compared to MM. For instance, it can incorporate application-specific biases into the optimization procedure without changing the objective function. We derive G-MM algorithms for several latent variable models and show empirically that they consistently outperform their MM counterparts in optimizing non-convex objectives. In particular, G-MM algorithms appear to be less sensitive to initialization. Keywords: majorization-minimization, non-convex optimization, latent variable models, expectation maximization 1. Introduction Non-convex optimization is ubiquitous in machine learning. Majorization-Minimization (MM) (Hunter et al., 2000) is an optimization framework for designing well-behaved optimization algorithms for non-convex functions. MM algorithms work by iteratively optimizing a sequence of easy-to-optimize surrogate functions that bound the objective. Two of the most successful instances of MM algorithms are Expectation-Maximization (EM) (Dempster et al., 1977) and the Concave-Convex Proce-1 arXiv:1506.07613v2 However, both have a number of drawbacks in practice, such as sensitivity to initialization and lack of uncertainty modeling for latent variables. This has been noted in works such as (Neal and Hinton, 1998; Felzenszwalb et al., 2010; Parizi et al., 2012; Kumar et al., 2012; Ping et al., 2014). We propose a new procedure, Generalized Majorization-Minimization (G-MM), for optimizing non-convex objective functions. Our approach is inspired by MM, but we generalize the bound construction process.

Massive MIMO is a variant of multiuser MIMO where the number of base-station antennas $M$ is very large (typically 100), and generally much larger than the number of spatially multiplexed data streams (typically 10). Unfortunately, the front-end A/D conversion necessary to drive hundreds of antennas, with a signal bandwidth of the order of 10 to 100 MHz, requires very large sampling bit-rate and power consumption. In order to reduce such implementation requirements, Hybrid Digital-Analog architectures have been proposed. In particular, our work in this paper is motivated by one of such schemes named Joint Spatial Division and Multiplexing (JSDM), where the downlink precoder (resp., uplink linear receiver) is split into the product of a baseband linear projection (digital) and an RF reconfigurable beamforming network (analog), such that only a reduced number $m \ll M$ of A/D converters and RF modulation/demodulation chains is needed. In JSDM, users are grouped according to the similarity of their channel dominant subspaces, and these groups are separated by the analog beamforming stage, where the multiplexing gain in each group is achieved using the digital precoder. Therefore, it is apparent that extracting the channel subspace information of the $M$-dim channel vectors from snapshots of $m$-dim projections, with $m \ll M$, plays a fundamental role in JSDM implementation. In this paper, we develop novel efficient algorithms that require sampling only $m = O(2\sqrt{M})$ specific array elements according to a coprime sampling scheme, and for a given $p \ll M$, return a $p$-dim beamformer that has a performance comparable with the best p-dim beamformer that can be designed from the full knowledge of the exact channel covariance matrix. We assess the performance of our proposed estimators both analytically and empirically via numerical simulations.

We consider clustering problems where the goal is to determine an optimal partition of a given point set in Euclidean space in terms of a collection of affine subspaces. While there is vast literature on heuristics for this kind of problem, such approaches are known to be susceptible to poor initializations and getting trapped in bad local optima. We alleviate these issues by introducing a semidefinite relaxation based on Lasserre's method of moments. While a similiar approach is known for classical Euclidean clustering problems, a generalization to our more general subspace scenario is not straightforward, due to the high symmetry of the objective function that weakens any convex relaxation. We therefore introduce a new mechanism for symmetry breaking based on covering the feasible region with polytopes. Additionally, we introduce and analyze a deterministic rounding heuristic.

We propose a scalable temporal latent space model for link prediction in dynamic social networks, where the goal is to predict links over time based on a sequence of previous graph snapshots. The model assumes that each user lies in an unobserved latent space and interactions are more likely to form between similar users in the latent space representation. In addition, the model allows each user to gradually move its position in the latent space as the network structure evolves over time. We present a global optimization algorithm to effectively infer the temporal latent space, with a quadratic convergence rate. Two alternative optimization algorithms with local and incremental updates are also proposed, allowing the model to scale to larger networks without compromising prediction accuracy. Empirically, we demonstrate that our model, when evaluated on a number of real-world dynamic networks, significantly outperforms existing approaches for temporal link prediction in terms of both scalability and predictive power.

Training machine learning models over massive amounts of data is a cornerstone of many data analytics tasks. Usually this involves solving large optimization problems involving millions of optimization variables and constraints. Doing so over a parallel platform, like Spark or Hadoop, is crucial to making such computations scalable. It is not always obvious how to solve large optimization problems in parallel. ADMM, which stands for the Alternating Directions Method of Multipliers, is a popular parallel optimization technique that provides a methodology for doing so.