Fine-tuning large pretrained language models on a limited training corpus usually suffers from poor generalization. Prior works show that the recently-proposed sharpness-aware minimization (SAM) optimization method can improve the model generalization. However, SAM adds a perturbation to each model parameter equally (but not all parameters contribute equally to the optimization of training), which we argue is sub-optimal and will lead to excessive computation. In this paper, we propose a novel optimization procedure, namely FSAM, which introduces a Fisher mask to improve the efficiency and performance of SAM. In short, instead of adding perturbation to all parameters, FSAM uses the Fisher information to identity the important parameters and formulates a Fisher mask to obtain the sparse perturbation, i.e., making the optimizer focus on these important parameters. Experiments on various tasks in GLUE and SuperGLUE benchmarks show that FSAM consistently outperforms the vanilla SAM by 0.67~1.98 average score among four different pretrained models. We also empirically show that FSAM works well in other complex scenarios, e.g., fine-tuning on generation tasks or limited training data. Encouragingly, when training data is limited, FSAM improves the SAM by a large margin, i.e., up to 15.1.

$\ell_0$ constrained optimization is prevalent in machine learning, particularly for high-dimensional problems, because it is a fundamental approach to achieve sparse learning. Hard-thresholding gradient descent is a dominant technique to solve this problem. However, first-order gradients of the objective function may be either unavailable or expensive to calculate in a lot of real-world problems, where zeroth-order (ZO) gradients could be a good surrogate. Unfortunately, whether ZO gradients can work with the hard-thresholding operator is still an unsolved problem. To solve this puzzle, in this paper, we focus on the $\ell_0$ constrained black-box stochastic optimization problems, and propose a new stochastic zeroth-order gradient hard-thresholding (SZOHT) algorithm with a general ZO gradient estimator powered by a novel random support sampling. We provide the convergence analysis of SZOHT under standard assumptions. Importantly, we reveal a conflict between the deviation of ZO estimators and the expansivity of the hard-thresholding operator, and provide a theoretical minimal value of the number of random directions in ZO gradients. In addition, we find that the query complexity of SZOHT is independent or weakly dependent on the dimensionality under different settings. Finally, we illustrate the utility of our method on a portfolio optimization problem as well as black-box adversarial attacks.

In this paper, we provide a novel strategy for solving Traveling Salesman Problem, which is a famous combinatorial optimization problem studied intensely in the TCS community. In particular, we consider the imitation learning framework, which helps a deterministic algorithm making good choices whenever it needs to, resulting in a speed up while maintaining the exactness of the solution without suffering from the unpredictability and a potential large deviation. Furthermore, we demonstrate a strong generalization ability of a graph neural network trained under the imitation learning framework. Specifically, the model is capable of solving a large instance of TSP faster than the baseline while has only seen small TSP instances when training.

In this paper we study the non-stationary stochastic optimization question with bandit feedback and dynamic regret measures. The seminal work of Besbes et al. (2015) shows that, when aggregated function changes is known a priori, a simple re-starting algorithm attains the optimal dynamic regret. In this work, we designed a stochastic optimization algorithm with fixed step sizes, which combined together with the multi-scale sampling framework of Wei and Luo (2021) achieves the optimal dynamic regret in non-stationary stochastic optimization without requiring prior knowledge of function change budget, thereby closes a question that has been open for a while. We also establish an additional result showing that any algorithm achieving good regret against stationary benchmarks with high probability could be automatically converted to an algorithm that achieves good regret against dynamic benchmarks, which is applicable to a wide class of bandit convex optimization algorithms.

We describe an augmented intelligence system for simplifying and enhancing the modeling experience for operations research. Using this system, the user receives a suggested formulation of an optimization problem based on its description. To facilitate this process, we build an intuitive user interface system that enables the users to validate and edit the suggestions. We investigate controlled generation techniques to obtain an automatic suggestion of formulation. Then, we evaluate their effectiveness with a newly created dataset of linear programming problems drawn from various application domains.

Multigraph matching is a recent variant of the graph matching problem. In this framework, the optimization procedure considers several graphs and enforces the consistency of the matches along the graphs. This constraint can be formalized as a cycle consistency across the pairwise permutation matrices, which implies the definition of a universe of vertex (Pachauri et al., 2013). The label of each vertex is encoded by a sparse vector and the dimension of this space corresponds to the rank of the bulk permutation matrix, the matrix built from the aggregation of all the pairwise permutation matrices. The matching problem can then be formulated as a non-convex quadratic optimization problem (QAP) under constraints imposed on the rank and the permutations. In this paper, we introduce a novel kernelized multigraph matching technique that handles vectors of attributes on both the vertices and edges of the graphs, while maintaining a low memory usage. We solve the QAP problem using a projected power optimization approach and propose several projectors leading to improved stability of the results. We provide several experiments showing that our method is competitive against other unsupervised methods.

In this work we consider a variant of adversarial online learning where in each round one picks $B$ out of $N$ arms and incurs cost equal to the $\textit{minimum}$ of the costs of each arm chosen. We propose an algorithm called Follow the Perturbed Multiple Leaders (FPML) for this problem, which we show (by adapting the techniques of Kalai and Vempala [2005]) achieves expected regret $\mathcal{O}(T^{\frac{1}{B+1}}\ln(N)^{\frac{B}{B+1}})$ over time horizon $T$ relative to the $\textit{single}$ best arm in hindsight. This introduces a trade-off between the budget $B$ and the single-best-arm regret, and we proceed to investigate several applications of this trade-off. First, we observe that algorithms which use standard regret minimizers as subroutines can sometimes be adapted by replacing these subroutines with FPML, and we use this to generalize existing algorithms for Online Submodular Function Maximization [Streeter and Golovin, 2008] in both the full feedback and semi-bandit feedback settings. Next, we empirically evaluate our new algorithms on an online black-box hyperparameter optimization problem. Finally, we show how FPML can lead to new algorithms for Linear Programming which require stronger oracles at the benefit of fewer oracle calls.

In this paper we consider finding an approximate second-order stationary point (SOSP) of nonconvex conic optimization that minimizes a twice differentiable function over the intersection of an affine subspace and a convex cone. In particular, we propose a Newton-conjugate gradient (Newton-CG) based barrier method for finding an $(\epsilon,\sqrt{\epsilon})$-SOSP of this problem. Our method is not only implementable, but also achieves an iteration complexity of ${\cal O}(\epsilon^{-3/2})$, which matches the best known iteration complexity of second-order methods for finding an $(\epsilon,\sqrt{\epsilon})$-SOSP of unconstrained nonconvex optimization. The operation complexity, consisting of ${\cal O}(\epsilon^{-3/2})$ Cholesky factorizations and $\widetilde{\cal O}(\epsilon^{-3/2}\min\{n,\epsilon^{-1/4}\})$ other fundamental operations, is also established for our method.

We design an model predictive control (MPC) approach for planning and control of non-holonomic mobile robots. Linearizing the system dynamics around the pre-computed reference trajectory gives a time-varying LQ MPC problem. We analytically show that by specially designing the MPC controller, the time-varying, linearized system can yield asymptotic stability around the origin in the tracking task. We further propose two obstacle avoidance methods. We show that by defining linearized constraint in velocity-space and explicitly coupling the two control inputs based on current state, our second method directly accounts for the non-holonomic property of the system and therefore alleviates infeasibility of the optimization problems. Simulation results suggest that regarding both static and dynamic obstacle avoidance, the planned trajectories by our LQ MPC approach are comparably smooth and effective as solving non-linear programming (NLP) problems, but in a more efficient way.

A Stackelberg congestion game (SCG) is a bilevel program in which a leader aims to maximize their own gain by anticipating and manipulating the equilibrium state at which followers settle by playing a congestion game. Large-scale SCGs are well known for their intractability and complexity. This study approaches SCGs through differentiable programming, which marries the latest developments in machine learning with conventional methodologies. The core idea centers on representing the lower-level equilibrium problem using an evolution path formed by the imitative logit dynamics. It enables the use of automatic differentiation over the evolution path towards equilibrium, leading to a double-loop gradient descent algorithm. We further show the fixation on the lower-level equilibrium may be a self-imposed computational obstacle. Instead, the leader may only look ahead along the followers' evolution path for a few steps, while updating their decisions in sync with the followers through a co-evolution process. The revelation gives rise to a single-loop algorithm that is more efficient in terms of both memory consumption and computation time. Through numerical experiments that cover a wide range of benchmark problems, we find the single-loop algorithm consistently strikes a good balance between solution quality and efficiency, outperforming not only the standard double-loop implementation but also other methods from the literature. Importantly, our results highlight both the wastefulness of "full anticipation" and the peril of "zero anticipation". If a quick-and-dirty heuristic is needed for solving a really large SCG, the proposed single-loop algorithm with a one-step look-ahead makes an ideal candidate.