In this talk, I present extensive evidence against the common belief that variational Bayesian learning is ineffective for large neural networks. I demonstrate that a direct optimization of the variational objective with an Improved Variational Online Newton method (IVON) can consistently match or outperforms Adam for training large networks such as GPT-2 and ResNets from scratch. IVON's computational costs are nearly identical to Adam but its predictive uncertainty is better. The talk concludes with several new use cases of variational learning where we improve fine-tuning and model merging in Large Language Models, accurately predict generalization error, and faithfully estimate sensitivity to data.

Global infrastructure robustness and local transport efficiency are critical requirements for transportation networks. However, since passengers often travel greedily to maximize their benefits and trigger traffic jams, transportation performance can be heavily disrupted. In this talk, we present a dynamical system that effectively routes passengers along their shortest paths while also strategically tuning the network edge weights to reduce congestion. As a result, we enforce both global and local optimality of transport. We show how the system is effective on synthetic networks and real-world data. Our findings on the international European highways suggest that thoughtfully devised routing schemes might help to lower car-produced carbon emissions.

We propose a new rule for adjusting the step size in the Stochastic Gradient method (SGD). This rule is related to that of the AdaGrad method but there are some significant differences. Most importantly, instead of using the norms of stochastic gradients, we use a stochastic approximation of the Bregman distance of the objective function. The resulting algorithm turns out to be the first universal method for Stochastic Optimization in the sense that it automatically adjusts not only to the oracle's noise level but also to the level of smoothness of the objective function. More specifically, our method has state-of-the-art worst-case convergence rate guarantees for the entire Hölder class of convex functions including both nonsmooth functions and those with Lipschitz continuous gradient. We also show how to use our approach for constructing an accelerated version of the Universal SGD with even better efficiency estimates.

We derive a moment relaxation for large-scale nonsmooth optimization problems with graphical structure and spherical constraints. In contrast to classical moment relaxations for global polynomial optimization that suffer from the curse of dimensionality we exploit the partially separable structure of the optimization problem to reduce the dimensionality of the search space. Leveraging optimal transport and Kantorovich-Rubinstein duality we decouple the problem and derive a tractable dual subspace approximation of the infinite-dimensional problem using spherical harmonics. This allows us to tackle polynomial optimization problems with spherical constraints and geodesic coupling terms. We show that the duality gap vanishes in the limit by proving that a Lipschitz continuous dual multiplier on a unit sphere can be approximated as closely as desired in terms of a Lipschitz continuous polynomial. The formulation is applied to sphere-valued imaging problems with total variation regularization and graph-based SLAM. In imaging tasks our approach achieves small duality gaps for a moderate degree. In graph-based SLAM our approach often finds solutions which after refinement with a local method are near the ground truth solution.

Motivated by the conspicuous use of momentum based algorithms in deep learning, we study a nonsmooth nonconvex stochastic heavy ball method and show its convergence. Our approach relies on semialgebraic assumptions, commonly met in practical situations, which allow to combine a conservative calculus with nonsmooth ODE methods. In particular, we can justify the use of subgradient sampling in practical implementations that employ backpropagation or implicit differentiation. Additionally, we provide general conditions for the sample distribution to ensure the convergence of the objective function. As for the stochastic subgradient method, our analysis highlights that subgradient sampling can make the stochastic heavy ball method converge to artificial critical points. We address this concern showing that these artifacts are almost surely avoided when initializations are randomized.

We study the variational inference problem of minimizing a regularized Rényi divergence over an exponential family, and propose a relaxed moment-matching algorithm, which includes a proximal-like step. Using the information-geometric link between Bregman divergences and the Kullback-Leibler divergence, this algorithm is shown to be equivalent to a Bregman proximal gradient algorithm. This novel perspective allows us to exploit the geometry of our approximate model while using stochastic black-box updates. We use this point of view to prove strong convergence guarantees including monotonic decrease of the objective, convergence to a stationary point or to the minimizer, and geometric convergence rates. These new theoretical insights lead to a versatile, robust, and competitive method, as illustrated by numerical experiments.

Gradient clipping is a popular modification to standard (stochastic) gradient descent, at every iteration limiting the gradient norm to a certain value c>0. It is widely used for example for stabilizing the training of deep learning models (Goodfellow et al., 2016), or for enforcing differential privacy (Abadi et al., 2016). Despite the popularity and the simplicity of the clipping mechanism, its convergence guarantees often require specific values of c and strong noise assumptions. In this talk, we discuss convergence guarantees that show precise dependence on arbitrary clipping thresholds c and show that our guarantees are tight with both deterministic and stochastic gradients.

We apply the PAC-Bayes theory to the setting of learning-to-optimize. To the best of our knowledge, we present the first framework to learn optimization algorithms with provable generalization guarantees (PAC-bounds) and explicit trade-off between a high probability of convergence and a high convergence speed. Even in the limit case, where convergence is guaranteed, our learned optimization algorithms provably outperform related algorithms based on a (deterministic) worst-case analysis. Our results rely on PAC-Bayes bounds for general, unbounded loss-functions based on exponential families. By generalizing existing ideas, we reformulate the learning procedure into a one-dimensional minimization problem and study the possibility to find a global minimum, which enables the algorithmic realization of the learning procedure. As a proof-of-concept, we learn hyperparameters of standard optimization algorithms to empirically underline our theory.

Fixed-Point Automatic Differentiation of Forward--Backward Splitting Algorithms for Partly Smooth Functions A large class of non-smooth practical optimization problems can be written as minimization of a sum of smooth and partly smooth functions. We consider such structured problems which also depend on a parameter vector and study the problem of differentiating its solution mapping with respect to the parameter which has far reaching applications in sensitivity analysis and parameter learning optmization problems. We show that under partial smoothness and other mild assumptions, Automatic Differentiation (AD) of the sequence generated by proximal splitting algorithms converges to the derivative of the solution mapping. For a variant of automatic differentiation, which we call Fixed-Point Automatic Differentiation (FPAD), we remedy the memory overhead problem of the Reverse Mode AD and moreover provide faster convergence theoretically. We numerically illustrate the convergence and convergence rates of AD and FPAD on Lasso and Group Lasso problems and demonstrate the working of FPAD on prototypical practical image denoising problem by learning the regularization term.

We introduce an inertial quasi-Newton Forward-Backward Splitting Algorithm to solve a class of monotone inclusion problems. While the inertial step is computationally cheap, in general, the bottleneck is the evaluation of the resolvent operator. A change of the metric makes its computation hard even for (otherwise in the standard metric) simple operators. In order to fully exploit the advantage of adapting the metric, we develop a new efficient resolvent calculus for a low-rank perturbed standard metric, which accounts exactly for quasi-Newton metrics. Moreover, we prove the convergence of our algorithms, including linear convergence rates in case one of the two considered operators is strongly monotone. Beyond the general monotone inclusion setup, we instantiate a novel inertial quasi-Newton Primal-Dual Hybrid Gradient Method for solving saddle point problems. The favourable performance of our inertial quasi-Newton PDHG method is demonstrated on several numerical experiments in image processing.

We focus on a stochastic minimization problem in the nonsmooth and nonconvex setting which applies for instance to the training of deep learning models. A popular way in the machine learning community to deal with this problem is to use stochastic gradient descent (SGD). This method combines both subgradient sampling and backpropagation, which is an efficient implementation of the chain-rule formula on nonsmooth compositions. Due to the incompatibility of these operations in the nonsmooth world, the SGD can generate spurious critical points in the optimization landscape which does not guarantee the convergence of the iterates to the critical set. We will explain in this talk how the model of Conservative Gradients is compatible with subgradient sampling and backpropagation, allowing to obtain convergence results for nonsmooth SGD. By means of definable geometry, we will emphasize that functions used in machine learning are locally endown with geometric properties of piecewise affine functions. In this setting, chain-ruling nonsmooth functions, and sampling subgradients output conservative gradients, but also, spurious critical points are hardly attained when performing SGD in practice.

We analyze the global and local behavior of gradient-like flows under stochastic errors towards the aim of solving convex optimization problems with noisy gradient input. We first study the unconstrained differentiable convex case, using a stochastic differential equation where the drift term is minus the gradient of the objective function and the diffusion term is either bounded or square-integrable. In this context, under Lipschitz continuity of the gradient, our first main result shows almost sure convergence of the objective and the trajectory process towards a minimizer of the objective function. We also provide a comprehensive complexity analysis by establishing several new pointwise and ergodic convergence rates in expectation for the convex, strongly convex, and (local) Lojasiewicz case. The latter, which involves local analysis, is challenging and requires non-trivial arguments from measure theory. Then, we extend our study to the constrained case and more generally to certain nonsmooth situations. We show that several of our results have natural extensions obtained by replacing the gradient of the objective function by a cocoercive monotone operator. This makes it possible to obtain similar convergence results for optimization problems with an additively "smooth + non-smooth" convex structure. Finally, we consider another extension of our results to non-smooth optimization which is based on the Moreau envelope.

We present a Lagrange decomposition method for solving 0-1 integer linear programs occurring in structured prediction. We propose a sequential and a massively min-marginal averaging schemes for solving the Lagrangean dual and a perturbation technique for decoding primal solutions. For representing subproblems we use binary decision diagrams (BDDs), which support efficient computation of subproblem solutions and update steps. We present experimental results on combinatorial problems from MAP inference for Markov Random Fields, quadratic assignment and cell tracking for developmental biology. Our highly parallel GPU implementation improves, comes close to, or outperform some state-of-the-art specialized heuristics while being problem agnostic.

Variational regularization techniques are dominant in the field of inverse problems. A drawback of these techniques is that they are dependent on a number of parameters which have to be set by the user. This issue can be approached by machine learning where we estimate these parameters from data. This is known as "Bilevel Learning" and has been successfully applied to many tasks, some as small-dimensional as learning a regularization parameter, others as high-dimensional as learning a sampling pattern in MRI. While mathematically appealing this strategy leads to a nested optimization problem which is computationally difficult to handle. In this talk we discuss several applications of bilevel learning for imaging as well as new computational approaches. There are quite a few open problems in this relatively recent field of study, some of which I will highlight along the way.

This presentation will go through several techniques for solving both scalar and vector optimization problems with various objective functions. On scalar version of optimization problems, we intend to discuss how third-order differential equations, or equations with an order higher than three, can contribute to development of fast optimization methods. The first problem to consider is to minimize the real-valued, convex, and continuously differentiable function defined on a Hilbert space. The problem will then be altered to minimize non- smooth and non-convex functions. We show that the convergence rate of the continuous dynamical system and the proximal algorithm obtained from discretization of it is the same. At the end of this route, we want to analyze the convergence of trajectories towards optimal solutions in both continuous and discrete versions. In the case of vector optimization problems, we aim to present a model of Alternating Direction Method of Multipliers (ADMM) which is a simple and powerful proximal algorithm, while the problem is minimization of sum of two vector-valued functions with a linear constraint respect to variables. This research is vital since there is a lack of literature on proximal techniques for tackling vector optimization problems.

Abstract: Min-max Optimization has broad applications in machine learning, e.g., Adversarial Robustness, AUC Maximization, Generative Adversarial Networks, etc. However, modern machine learning models are intrinsically nonconvex (e.g., deep learning, matrix factorization), which brings tremendous challenge for min-max optimization. In this talk, I will talk about my work on provably efficient algorithms for min-max optimization in the presence of non-convexity. This talk will showcase a few results. First, I will talk about nonconvex-concave min-max optimization, with applications in deep AUC optimization. We designed polynomial-time algorithms to converge to stationary point, as well as maximal AUC value, under different assumptions. Second, I will talk about nonconvex-nonconcave min-max optimization, with applications in training Generative Adversarial Networks. The key result is that under the Minty Variational Inequality (MVI) condition, we design polynomial-time algorithms to find stationary points.

In this talk, we present SIRTR (Stochastic Inexact Restoration Trust-Region method) for solving finite-sum minimization problems. At each iteration, SIRTR approximates both function and gradient by sampling. The function sample size is computed using a deterministic rule inspired by the inexact restoration method, whereas the gradient sample size can be smaller than the sample size employed in function approximations. Notably, our approach may allow the decrease of the sample sizes at some iterations. We show that SIRTR eventually reaches full precision in evaluating the objective function and we provide a worst-case complexity result on the number of iterations required to achieve full precision. Numerical results on nonconvex binary classification problems confirm that SIRTR is able to provide accurate approximations way before the maximum sample size is reached and without requiring a problem-dependent tuning of the parameters involved.

Abstract: First-order operator splitting methods are ubiquitous among many fields through science and engineering, such as inverse problems, image processing, statistics, data science and machine learning, to name a few. In this talk, through the fixed-point sequence, I will first discuss a geometry property of first-order methods when applying to solve non-smooth optimization problems. Then I will discuss the limitation of current widely used "inertial acceleration" technique, and propose a trajectory following adaptive acceleration algorithm. Global convergence is established for the proposed acceleration scheme based on the perturbation of fixed-point iteration. Locally, connections between the acceleration scheme and the well studied "vector extrapolation technique" in the field of numerical analysis will be discussed, followed by acceleration guarantees of the proposed acceleration scheme. Numeric experiments on various first-order methods are provided to demonstrate the advantage of the proposed adaptive acceleration scheme.

Abstract: Finding the optimal hyperparameters of a model can be cast as a bilevel optimization problem, typically solved using zero-order techniques. In this work we study first-order methods when the inner optimization problem is convex but non-smooth. We show that the forward-mode differentiation of proximal gradient descent and proximal coordinate descent yield sequences of Jacobians converging toward the exact Jacobian. Using implicit differentiation, we show it is possible to leverage the non-smoothness of the inner problem to speed up the computation. Finally, we provide a bound on the error made on the hypergradient when the inner optimization problem is solved approximately. Results on regression and classification problems reveal computational benefits for hyperparameter optimization, especially when multiple hyperparameters are required.

Abstract: When solving multiple optimization problems sharing the same underlying structure, using iterative algorithms designed for worst case scenario can be considered as inefficient. When one aim at having good solution in average, it is possible to improve the performances by learning the weights of a neural networked designed to mimic an unfolded optimization algorithm. However, the reason why learning the weights of such a network would accelerate the problem resolution is not always clear. In this talk, I will first review how one can design unrolled algorithms to solve the linear regression with l1 or TV regularization, with a particular focus on the choice of parametrization and loss. Then, I will discuss the reasons why such procedure can lead to better results compared to classical optimization, with a particular focus on the choice of step sizes.