Search

We consider the problem of optimizing the sum of a smooth, nonconvex function for which derivatives are unavailable, and a convex, nonsmooth function with easy-to-evaluate proximal operator. Of particular focus is the case where the smooth part has a nonlinear least-squares structure. We adapt two existing approaches for derivative-free optimization of nonsmooth compositions of smooth functions to this setting. Our main contribution is adapting our algorithm to handle inexactly computed stationary measures, where the inexactness is adaptively adjusted as required by the algorithm (where previous approaches assumed access to exact stationary measures, which is not realistic in this setting). Numerically, we provide two extensions of the state-of-the-art DFO-LS solver for nonlinear least-squares problems and demonstrate their strong practical performance.

We develop a new approximation theory for linear and quadratic interpolation models, suitable for use in convex-constrained derivative-free optimization (DFO). Most existing model-based DFO methods for constrained problems assume the ability to construct sufficiently accurate approximations via interpolation, but the standard notions of accuracy (designed for unconstrained problems) may not be achievable by only sampling feasible points, and so may not give practical algorithms. This work extends the theory of convex-constrained linear interpolation developed in [Hough & Roberts, SIAM J. Optim, 32:4 (2022), pp. 2552-2579] to the case of linear regression models and underdetermined quadratic interpolation models.

The analysis of gradient descent-type methods typically relies on the Lipschitz continuity of the objective gradient. This generally requires an expensive hyperparameter tuning process to appropriately calibrate a stepsize for a given problem. In this work we introduce a local first-order smoothness oracle (LFSO) which generalizes the Lipschitz continuous gradients smoothness condition and is applicable to any twice-differentiable function. We show that this oracle can encode all relevant problem information for tuning stepsizes for a suitably modified gradient descent method and give global and local convergence results. We also show that LFSOs in this modified first-order method can yield global linear convergence rates for non-strongly convex problems with extremely flat minima, and thus improve over the lower bound on rates achievable by general (accelerated) first-order methods.

Various tasks in data science are modeled utilizing the variational regularization approach, where manually selecting regularization parameters presents a challenge. The difficulty gets exacerbated when employing regularizers involving a large number of hyperparameters. To overcome this challenge, bilevel learning can be employed to learn such parameters from data. However, neither exact function values nor exact gradients with respect to the hyperparameters are attainable, necessitating methods that only rely on inexact evaluation of such quantities. State-of-the-art inexact gradient-based methods a priori select a sequence of the required accuracies and cannot identify an appropriate step size since the Lipschitz constant of the hypergradient is unknown. In this work, we propose an algorithm with backtracking line search that only relies on inexact function evaluations and hypergradients and show convergence to a stationary point. Furthermore, the proposed algorithm determines the required accuracy dynamically rather than manually selected before running it. Our numerical experiments demonstrate the efficiency and feasibility of our approach for hyperparameter estimation on a range of relevant problems in imaging and data science such as total variation and field of experts denoising and multinomial logistic regression. Particularly, the results show that the algorithm is robust to its own hyperparameters such as the initial accuracies and step size.

Derivative-free algorithms seek the minimum of a given function based only on function values queried at appropriate points. Although these methods are widely used in practice, their performance is known to worsen as the problem dimension increases. Recent advances in developing randomized derivative-free techniques have tackled this issue by working in low-dimensional subspaces that are drawn at random in an iterative fashion. The connection between the dimension of these random subspaces and the algorithmic guarantees has yet to be fully understood. In this paper, we develop an analysis for derivative-free algorithms (both direct-search and model-based approaches) employing random subspaces. Our results leverage linear local approximations of smooth functions to obtain understanding of the expected decrease achieved per function evaluation. Although the quantities of interest involve multidimensional integrals with no closed-form expression, a relative comparison for different subspace dimensions suggest that low dimension is preferable. Numerical computation of the quantities of interest confirm the benefit of operating in low-dimensional subspaces.

Estimating hyperparameters has been a long-standing problem in machine learning. We consider the case where the task at hand is modeled as the solution to an optimization problem. Here the exact gradient with respect to the hyperparameters cannot be feasibly computed and approximate strategies are required. We introduce a unified framework for computing hypergradients that generalizes existing methods based on the implicit function theorem and automatic differentiation/backpropagation, showing that these two seemingly disparate approaches are actually tightly connected. Our framework is extremely flexible, allowing its subproblems to be solved with any suitable method, to any degree of accuracy. We derive a priori and computable a posteriori error bounds for all our methods, and numerically show that our a posteriori bounds are usually more accurate. Our numerical results also show that, surprisingly, for efficient bilevel optimization, the choice of hypergradient algorithm is at least as important as the choice of lower-level solver., Comment: Accepted to IMA Journal of Applied Mathematics

Classical ghost imaging is a new paradigm in imaging where the image of an object is not measured directly with a pixelated detector. Rather, the object is subject to a set of illumination patterns and the total interaction of the object, e.g., reflected or transmitted photons or particles, is measured for each pattern with a single-pixel or bucket detector. An image of the object is then computed through the correlation of each pattern and the corresponding bucket value. Assuming no prior knowledge of the object, the set of patterns used to compute the ghost image dictates the image quality. In the visible-light regime, programmable spatial light modulators can generate the illumination patterns. In many other regimes -- such as x rays, electrons, and neutrons -- no such dynamically configurable modulators exist, and patterns are commonly produced by employing a transversely-translated mask. In this paper we explore some of the properties of masks or speckle that should be considered to maximize ghost-image quality, given a certain experimental classical ghost-imaging setup employing a transversely-displaced but otherwise non-configurable mask., Comment: 28 pages, 17 figures

In distributed learning, a central server trains a model according to updates provided by nodes holding local data samples. In the presence of one or more malicious servers sending incorrect information (a Byzantine adversary), standard algorithms for model training such as stochastic gradient descent (SGD) fail to converge. In this paper, we present a simplified convergence theory for the generic Byzantine Resilient SGD method originally proposed by Blanchard et al. [NeurIPS 2017]. Compared to the existing analysis, we shown convergence to a stationary point in expectation under standard assumptions on the (possibly nonconvex) objective function and flexible assumptions on the stochastic gradients.

In this paper, we study a generic direct-search algorithm in which the polling directions are defined using random subspaces. Complexity guarantees for such an approach are derived thanks to probabilistic properties related to both the subspaces and the directions used within these subspaces. Our analysis crucially extends previous deterministic and probabilistic arguments by relaxing the need for directions to be deterministically bounded in norm. As a result, our approach encompasses a wide range of new optimal polling strategies that can be characterized using our subspace and direction properties. By leveraging results on random subspace embeddings and sketching matrices, we show that better complexity bounds are obtained for randomized instances of our framework. A numerical investigation confirms the benefit of randomization, particularly when done in subspaces, when solving problems of moderately large dimension.

We present a model-based derivative-free method for optimization subject to general convex constraints, which we assume are unrelaxable and accessed only through a projection operator that is cheap to evaluate. We prove global convergence and a worst-case complexity of $O(\epsilon^{-2})$ iterations and objective evaluations for nonconvex functions, matching results for the unconstrained case. We introduce new, weaker requirements on model accuracy compared to existing methods. As a result, sufficiently accurate interpolation models can be constructed only using feasible points. We develop a comprehensive theory of interpolation set management in this regime for linear and composite linear models. We implement our approach for nonlinear least-squares problems and demonstrate strong practical performance compared to general-purpose solvers.