A Symplectic Analysis of Alternating Mirror Descent

Motivated by understanding the behavior of the Alternating Mirror Descent (AMD) algorithm for bilinear zero-sum games, we study the discretization of continuous-time Hamiltonian flow via the symplectic Euler method. We provide a framework for analysis using results from Hamiltonian dynamics, Lie algebra, and symplectic numerical integrators, with an emphasis on the existence and properties of a conserved quantity, the modified Hamiltonian (MH), for the symplectic Euler method. We compute the MH in closed-form when the original Hamiltonian is a quadratic function, and show that it generally differs from the other conserved quantity known previously in that case. We derive new error bounds on the MH when truncated at orders in the stepsize in terms of the number of iterations, $K$, and utilize this bound to show an improved $\mathcal{O}(K^{1/5})$ total regret bound and an $\mathcal{O}(K^{-4/5})$ duality gap of the average iterates for AMD. Finally, we propose a conjecture which, if true, would imply that the total regret for AMD goes as $\mathcal{O}\left(K^{\varepsilon}\right)$ and the duality gap of the average iterates as $\mathcal{O}\left(K^{-1+\varepsilon}\right)$ for any $\varepsilon>0$, and we can take $\varepsilon=0$ upon certain convergence conditions for the MH.

On Independent Samples Along the Langevin Diffusion and the Unadjusted Langevin Algorithm

We study the rate at which the initial and current random variables become independent along a Markov chain, focusing on the Langevin diffusion in continuous time and the Unadjusted Langevin Algorithm (ULA) in discrete time. We measure the dependence between random variables via their mutual information. For the Langevin diffusion, we show the mutual information converges to $0$ exponentially fast when the target is strongly log-concave, and at a polynomial rate when the target is weakly log-concave. These rates are analogous to the mixing time of the Langevin diffusion under similar assumptions. For the ULA, we show the mutual information converges to $0$ exponentially fast when the target is strongly log-concave and smooth. We prove our results by developing the mutual version of the mixing time analyses of these Markov chains. We also provide alternative proofs based on strong data processing inequalities for the Langevin diffusion and the ULA, and by showing regularity results for these processes in mutual information.

Optimal score estimation via empirical Bayes smoothing

We study the problem of estimating the score function of an unknown probability distribution $\rho^*$ from $n$ independent and identically distributed observations in $d$ dimensions. Assuming that $\rho^*$ is subgaussian and has a Lipschitz-continuous score function $s^*$, we establish the optimal rate of $\tilde \Theta(n^{-\frac{2}{d+4}})$ for this estimation problem under the loss function $\|\hat s - s^*\|^2_{L^2(\rho^*)}$ that is commonly used in the score matching literature, highlighting the curse of dimensionality where sample complexity for accurate score estimation grows exponentially with the dimension $d$. Leveraging key insights in empirical Bayes theory as well as a new convergence rate of smoothed empirical distribution in Hellinger distance, we show that a regularized score estimator based on a Gaussian kernel attains this rate, shown optimal by a matching minimax lower bound. We also discuss the implication of our theory on the sample complexity of score-based generative models.

Fast sampling from constrained spaces using the Metropolis-adjusted Mirror Langevin Algorithm

We propose a new method called the Metropolis-adjusted Mirror Langevin algorithm for approximate sampling from distributions whose support is a compact and convex set. This algorithm adds an accept-reject filter to the Markov chain induced by a single step of the mirror Langevin algorithm (Zhang et al., 2020), which is a basic discretisation of the mirror Langevin dynamics. Due to the inclusion of this filter, our method is unbiased relative to the target, while known discretisations of the mirror Langevin dynamics including the mirror Langevin algorithm have an asymptotic bias. We give upper bounds for the mixing time of the proposed algorithm when the potential is relatively smooth, convex, and Lipschitz with respect to a self-concordant mirror function. As a consequence of the reversibility of the Markov chain induced by the algorithm, we obtain an exponentially better dependence on the error tolerance for approximate sampling. We also present numerical experiments that corroborate our theoretical findings.

Extragradient Type Methods for Riemannian Variational Inequality Problems

Riemannian convex optimization and minimax optimization have recently drawn considerable attention. Their appeal lies in their capacity to adeptly manage the non-convexity of the objective function as well as constraints inherent in the feasible set in the Euclidean sense. In this work, we delve into monotone Riemannian Variational Inequality Problems (RVIPs), which encompass both Riemannian convex optimization and minimax optimization as particular cases. In the context of Euclidean space, it is established that the last-iterates of both the extragradient (EG) and past extragradient (PEG) methods converge to the solution of monotone variational inequality problems at a rate of $O\left(\frac{1}{\sqrt{T}}\right)$ (Cai et al., 2022). However, analogous behavior on Riemannian manifolds remains an open question. To bridge this gap, we introduce the Riemannian extragradient (REG) and Riemannian past extragradient (RPEG) methods. We demonstrate that both exhibit $O\left(\frac{1}{\sqrt{T}}\right)$ last-iterate convergence. Additionally, we show that the average-iterate convergence of both REG and RPEG is $O\left(\frac{1}{{T}}\right)$, aligning with observations in the Euclidean case (Mokhtari et al., 2020). These results are enabled by judiciously addressing the holonomy effect so that additional complications in Riemannian cases can be reduced and the Euclidean proof inspired by the performance estimation problem (PEP) technique or the sum-of-squares (SOS) technique can be applied again.

Mitigating Catastrophic Forgetting in Long Short-Term Memory Networks

Continual learning on sequential data is critical for many machine learning (ML) deployments. Unfortunately, LSTM networks, which are commonly used to learn on sequential data, suffer from catastrophic forgetting and are limited in their ability to learn multiple tasks continually. We discover that catastrophic forgetting in LSTM networks can be overcome in two novel and readily-implementable ways -- separating the LSTM memory either for each task or for each target label. Our approach eschews the need for explicit regularization, hypernetworks, and other complex methods. We quantify the benefits of our approach on recently-proposed LSTM networks for computer memory access prefetching, an important sequential learning problem in ML-based computer system optimization. Compared to state-of-the-art weight regularization methods to mitigate catastrophic forgetting, our approach is simple, effective, and enables faster learning. We also show that our proposal enables the use of small, non-regularized LSTM networks for complex natural language processing in the offline learning scenario, which was previously considered difficult.

Continuized Acceleration for Quasar Convex Functions in Non-Convex Optimization

Quasar convexity is a condition that allows some first-order methods to efficiently minimize a function even when the optimization landscape is non-convex. Previous works develop near-optimal accelerated algorithms for minimizing this class of functions, however, they require a subroutine of binary search which results in multiple calls to gradient evaluations in each iteration, and consequently the total number of gradient evaluations does not match a known lower bound. In this work, we show that a recently proposed continuized Nesterov acceleration can be applied to minimizing quasar convex functions and achieves the optimal bound with a high probability. Furthermore, we find that the objective functions of training generalized linear models (GLMs) satisfy quasar convexity, which broadens the applicability of the relevant algorithms, while known practical examples of quasar convexity in non-convex learning are sparse in the literature. We also show that if a smooth and one-point strongly convex, Polyak-Lojasiewicz, or quadratic-growth function satisfies quasar convexity, then attaining an accelerated linear rate for minimizing the function is possible under certain conditions, while acceleration is not known in general for these classes of functions.

Convergence in KL Divergence of the Inexact Langevin Algorithm with Application to Score-based Generative Models

We study the Inexact Langevin Algorithm (ILA) for sampling using estimated score function when the target distribution satisfies log-Sobolev inequality (LSI), motivated by Score-based Generative Modeling (SGM). We prove a long-term convergence in Kullback-Leibler (KL) divergence under a sufficient assumption that the error of the score estimator has a bounded Moment Generating Function (MGF). Our assumption is weaker than $L^\infty$ (which is too strong to hold in practice) and stronger than $L^2$ error assumption, which we show not sufficient to guarantee convergence in general. Under the $L^\infty$ error assumption, we additionally prove convergence in R\'enyi divergence, which is stronger than KL divergence. We then study how to get a provably accurate score estimator which satisfies bounded MGF assumption for LSI target distributions, by using an estimator based on kernel density estimation. Together with the convergence results, we yield the first end-to-end convergence guarantee for ILA in the population level. Last, we generalize our convergence analysis to SGM and derive a complexity guarantee in KL divergence for data satisfying LSI under MGF-accurate score estimator.

Aggregation in the Mirror Space : Fast, Accurate Distributed Machine Learning in Military Settings

Distributed machine learning (DML) can be an important capability for modern military to take advantage of data and devices distributed at multiple vantage points to adapt and learn. The existing distributed machine learning frameworks, however, cannot realize the full benefits of DML, because they are all based on the simple linear aggregation framework, but linear aggregation cannot handle the $\textit{divergence challenges}$ arising in military settings: the learning data at different devices can be heterogeneous ($\textit{i.e.}$, Non-IID data), leading to model divergence, but the ability for devices to communicate is substantially limited ($\textit{i.e.}$, weak connectivity due to sparse and dynamic communications), reducing the ability for devices to reconcile model divergence. In this paper, we introduce a novel DML framework called aggregation in the mirror space (AIMS) that allows a DML system to introduce a general mirror function to map a model into a mirror space to conduct aggregation and gradient descent. Adapting the convexity of the mirror function according to the divergence force, AIMS allows automatic optimization of DML. We conduct both rigorous analysis and extensive experimental evaluations to demonstrate the benefits of AIMS. For example, we prove that AIMS achieves a loss of $O\left((\frac{m^{r+1}}{T})^{\frac1r}\right)$ after $T$ network-wide updates, where $m$ is the number of devices and $r$ the convexity of the mirror function, with existing linear aggregation frameworks being a special case with $r=2$. Our experimental evaluations using EMANE (Extendable Mobile Ad-hoc Network Emulator) for military communications settings show similar results: AIMS can improve DML convergence rate by up to 57\% and scale well to more devices with weak connectivity, all with little additional computation overhead compared to traditional linear aggregation.

Towards Understanding GD with Hard and Conjugate Pseudo-labels for Test-Time Adaptation

We consider a setting that a model needs to adapt to a new domain under distribution shifts, given that only unlabeled test samples from the new domain are accessible at test time. A common idea in most of the related works is constructing pseudo-labels for the unlabeled test samples and applying gradient descent (GD) to a loss function with the pseudo-labels. Recently, Goyal et al. (2022) propose conjugate labels, which is a new kind of pseudo-labels for self-training at test time. They empirically show that the conjugate label outperforms other ways of pseudo-labeling on many domain adaptation benchmarks. However, provably showing that GD with conjugate labels learns a good classifier for test-time adaptation remains open. In this work, we aim at theoretically understanding GD with hard and conjugate labels for a binary classification problem. We show that for square loss, GD with conjugate labels converges to a solution that minimizes the testing 0-1 loss under a Gaussian model, while GD with hard pseudo-labels fails in this task. We also analyze them under different loss functions for the update. Our results shed lights on understanding when and why GD with hard labels or conjugate labels works in test-time adaptation.