Gradient descent-ascent (GDA) is a widely used algorithm for minimax optimization. However, GDA has been proved to converge to stationary points for nonconvex minimax optimization, which are suboptimal compared with local minimax points. In this work, we develop cubic regularization (CR) type algorithms that globally converge to local minimax points in nonconvex-strongly-concave minimax optimization. We first show that local minimax points are equivalent to second-order stationary points of a certain envelope function. Then, inspired by the classic cubic regularization algorithm, we propose an algorithm named Cubic-LocalMinimax for finding local minimax points, and provide a comprehensive convergence analysis by leveraging its intrinsic potential function. Specifically, we establish the global convergence of Cubic-LocalMinimax to a local minimax point at a sublinear convergence rate and characterize its iteration complexity. Also, we propose a GDA-based solver for solving the cubic subproblem involved in Cubic-LocalMinimax up to certain pre-defined accuracy, and analyze the overall gradient and Hessian-vector product computation complexities of such an inexact Cubic-LocalMinimax algorithm. Moreover, we propose a stochastic variant of Cubic-LocalMinimax for large-scale minimax optimization, and characterize its sample complexity under stochastic sub-sampling. Experimental results demonstrate faster convergence of our stochastic Cubic-LocalMinimax than some existing algorithms.

Stochastic gradient descent is a workhorse for training deep neural networks due to its excellent generalization performance. Several studies demonstrated this success is attributed to the implicit bias of the method that prefers a flat minimum and developed new methods based on this perspective. Recently, Izmailov et al. (2018) empirically observed that an averaged stochastic gradient descent with a large step size can bring out the implicit bias more effectively and can converge more stably to a flat minimum than the vanilla stochastic gradient descent. In our work, we theoretically justify this observation by showing that the averaging scheme improves the bias-optimization tradeoff coming from the stochastic gradient noise: a large step size amplifies the bias but makes convergence unstable, and vice versa. Specifically, we show that the averaged stochastic gradient descent can get closer to a solution of a penalized objective on the sharpness than the vanilla stochastic gradient descent using the same step size under certain conditions. In experiments, we verify our theory and show this learning scheme significantly improves performance.

Despite being highly over-parametrized, and having the ability to fully interpolate the training data, deep networks are known to generalize well to unseen data. It is now understood that part of the reason for this is that the training algorithms used have certain implicit regularization properties that ensure interpolating solutions with "good" properties are found. This is best understood in linear over-parametrized models where it has been shown that the celebrated stochastic gradient descent (SGD) algorithm finds an interpolating solution that is closest in Euclidean distance to the initial weight vector. Different regularizers, replacing Euclidean distance with Bregman divergence, can be obtained if we replace SGD with stochastic mirror descent (SMD). Empirical observations have shown that in the deep network setting, SMD achieves a generalization performance that is different from that of SGD (and which depends on the choice of SMD's potential function. In an attempt to begin to understand this behavior, we obtain the generalization error of SMD for over-parametrized linear models for a binary classification problem where the two classes are drawn from a Gaussian mixture model. We present simulation results that validate the theory and, in particular, introduce two data models, one for which SMD with an $\ell_2$ regularizer (i.e., SGD) outperforms SMD with an $\ell_1$ regularizer, and one for which the reverse happens.

We develop an algorithmic framework for solving convex optimization problems using no-regret game dynamics. By converting the problem of minimizing a convex function into an auxiliary problem of solving a min-max game in a sequential fashion, we can consider a range of strategies for each of the two-players who must select their actions one after the other. A common choice for these strategies are so-called no-regret learning algorithms, and we describe a number of such and prove bounds on their regret. We then show that many classical first-order methods for convex optimization -- including average-iterate gradient descent, the Frank-Wolfe algorithm, Nesterov's acceleration methods, and the accelerated proximal method -- can be interpreted as special cases of our framework as long as each player makes the correct choice of no-regret strategy. Proving convergence rates in this framework becomes very straightforward, as they follow from plugging in the appropriate known regret bounds. Our framework also gives rise to a number of new first-order methods for special cases of convex optimization that were not previously known.

As privacy protection receives much attention, unlearning the effect of a specific node from a pre-trained graph learning model has become equally important. However, due to the node dependency in the graph-structured data, representation unlearning in Graph Neural Networks (GNNs) is challenging and less well explored. In this paper, we fill in this gap by first studying the unlearning problem in linear-GNNs, and then introducing its extension to non-linear structures. Given a set of nodes to unlearn, we propose PROJECTOR that unlearns by projecting the weight parameters of the pre-trained model onto a subspace that is irrelevant to features of the nodes to be forgotten. PROJECTOR could overcome the challenges caused by node dependency and enjoys a perfect data removal, i.e., the unlearned model parameters do not contain any information about the unlearned node features which is guaranteed by algorithmic construction. Empirical results on real-world datasets illustrate the effectiveness and efficiency of PROJECTOR.

Federated Learning (FL) is a nascent decentralized learning framework under which a massive collection of heterogeneous clients collaboratively train a model without revealing their local data. Scarce communication, privacy leakage, and Byzantine attacks are the key bottlenecks of system scalability. In this paper, we focus on communication-efficient distributed (stochastic) gradient descent for non-convex optimization, a driving force of FL. We propose two algorithms, named {\em Adaptive Stochastic Sign SGD (Ada-StoSign)} and {\em $\beta$-Stochastic Sign SGD ($\beta$-StoSign)}, each of which compresses the local gradients into bit vectors. To handle unbounded gradients, Ada-StoSign uses a novel norm tracking function that adaptively adjusts a coarse estimation on the $\ell_{\infty}$ of the local gradients - a key parameter used in gradient compression. We show that Ada-StoSign converges in expectation with a rate $O(\log T/\sqrt{T} + 1/\sqrt{M})$, where $M$ is the number of clients. To the best of our knowledge, when $M$ is sufficiently large, Ada-StoSign outperforms the state-of-the-art sign-based method whose convergence rate is $O(T^{-1/4})$. Under bounded gradient assumption, $\beta$-StoSign achieves quantifiable Byzantine resilience and privacy assurances, and works with partial client participation and mini-batch gradients which could be unbounded. We corroborate and complement our theories by experiments on MNIST and CIFAR-10 datasets.

When training neural networks, it has been widely observed that a large step size is essential in stochastic gradient descent (SGD) for obtaining superior models. However, the effect of large step sizes on the success of SGD is not well understood theoretically. Several previous works have attributed this success to the stochastic noise present in SGD. However, we show through a novel set of experiments that the stochastic noise is not sufficient to explain good non-convex training, and that instead the effect of a large learning rate itself is essential for obtaining best performance.We demonstrate the same effects also in the noise-less case, i.e. for full-batch GD. We formally prove that GD with large step size -- on certain non-convex function classes -- follows a different trajectory than GD with a small step size, which can lead to convergence to a global minimum instead of a local one. Our settings provide a framework for future analysis which allows comparing algorithms based on behaviors that can not be observed in the traditional settings.

The Sharpness Aware Minimization (SAM) optimization algorithm has been shown to control large eigenvalues of the loss Hessian and provide generalization benefits in a variety of settings. The original motivation for SAM was a modified loss function which penalized sharp minima; subsequent analyses have also focused on the behavior near minima. However, our work reveals that SAM provides a strong regularization of the eigenvalues throughout the learning trajectory. We show that in a simplified setting, SAM dynamically induces a stabilization related to the edge of stability (EOS) phenomenon observed in large learning rate gradient descent. Our theory predicts the largest eigenvalue as a function of the learning rate and SAM radius parameters. Finally, we show that practical models can also exhibit this EOS stabilization, and that understanding SAM must account for these dynamics far away from any minima.

We prove that various stochastic gradient descent methods, including the stochastic gradient descent (SGD), stochastic heavy-ball (SHB), and stochastic Nesterov's accelerated gradient (SNAG) methods, almost surely avoid any strict saddle manifold. To the best of our knowledge, this is the first time such results are obtained for SHB and SNAG methods. Moreover, our analysis expands upon previous studies on SGD by removing the need for bounded gradients of the objective function and uniformly bounded noise. Instead, we introduce a more practical local boundedness assumption for the noisy gradient, which is naturally satisfied in empirical risk minimization problems typically seen in training of neural networks. Keywords: Stochastic gradient descent, stochastic heavy-ball, stochastic Nesterov's accelerated gradient, almost sure saddle avoidance

We propose new limiting dynamics for stochastic gradient descent in the small learning rate regime called stochastic modified flows. These SDEs are driven by a cylindrical Brownian motion and improve the so-called stochastic modified equations by having regular diffusion coefficients and by matching the multi-point statistics. As a second contribution, we introduce distribution dependent stochastic modified flows which we prove to describe the fluctuating limiting dynamics of stochastic gradient descent in the small learning rate - infinite width scaling regime.