arXiv subjects

John C. Duchi

Publications and source records attributed to John C. Duchi.

At least 37 records · Page 2Linked to original sources

arXiv2020

The Right Complexity Measure in Locally Private Estimation: It is not the Fisher Information

We identify fundamental tradeoffs between statistical utility and privacy under local models of privacy in which data is kept private even from the statistician, providing instance-specific bounds for private estimation and learning problems by developing the \emph{local minimax risk}. In contrast to approaches based on worst-case (minimax) error, which are conservative, this allows us to evaluate the difficulty of individual problem instances and delineate the possibilities for adaptation in private estimation and inference. Our main results show that the local modulus of continuity of the estimand with respect to the variation distance---as opposed to the Hellinger distance central to classical statistics---characterizes rates of convergence under locally private estimation for many notions of privacy, including differential privacy and its relaxations. As consequences of these results, we identify an alternative to the Fisher information for private estimation, giving a more nuanced understanding of the challenges of adaptivity and optimality.

math.ST↗

arXiv2020

Second-Order Information in Non-Convex Stochastic Optimization: Power and Limitations

We design an algorithm which finds an $ε$-approximate stationary point (with $\|\nabla F(x)\|\le ε$) using $O(ε^{-3})$ stochastic gradient and Hessian-vector products, matching guarantees that were previously available only under a stronger assumption of access to multiple queries with the same random seed. We prove a lower bound which establishes that this rate is optimal and---surprisingly---that it cannot be improved using stochastic $p$th order methods for any $p\ge 2$, even when the first $p$ derivatives of the objective are Lipschitz. Together, these results characterize the complexity of non-convex stochastic optimization with second-order methods and beyond. Expanding our scope to the oracle complexity of finding $(ε,γ)$-approximate second-order stationary points, we establish nearly matching upper and lower bounds for stochastic second-order methods. Our lower bounds here are novel even in the noiseless case.

cs.LG↗

arXiv2020

Near Instance-Optimality in Differential Privacy

We develop two notions of instance optimality in differential privacy, inspired by classical statistical theory: one by defining a local minimax risk and the other by considering unbiased mechanisms and analogizing the Cramer-Rao bound, and we show that the local modulus of continuity of the estimand of interest completely determines these quantities. We also develop a complementary collection mechanisms, which we term the inverse sensitivity mechanisms, which are instance optimal (or nearly instance optimal) for a large class of estimands. Moreover, these mechanisms uniformly outperform the smooth sensitivity framework on each instance for several function classes of interest, including real-valued continuous functions. We carefully present two instantiations of the mechanisms for median and robust regression estimation with corresponding experiments.

cs.CR↗

arXiv2020

A constrained risk inequality for general losses

We provide a general constrained risk inequality that applies to arbitrary non-decreasing losses, extending a result of Brown and Low [Ann. Stat. 1996]. Given two distributions $P_0$ and $P_1$, we find a lower bound for the risk of estimating a parameter $θ(P_1)$ under $P_1$ given an upper bound on the risk of estimating the parameter $θ(P_0)$ under $P_0$. The inequality is a useful pedagogical tool, as its proof relies only on the Cauchy-Schwartz inequality, it applies to general losses, and it transparently gives risk lower bounds on super-efficient and adaptive estimators.

math.ST↗

arXiv2020

First-Order Methods for Nonconvex Quadratic Minimization

We consider minimization of indefinite quadratics with either trust-region (norm) constraints or cubic regularization. Despite the nonconvexity of these problems we prove that, under mild assumptions, gradient descent converges to their global solutions, and give a non-asymptotic rate of convergence for the cubic variant. We also consider Krylov subspace solutions and establish sharp convergence guarantees to the solutions of both trust-region and cubic-regularized problems. Our rates mirror the behavior of these methods on convex quadratics and eigenvector problems, highlighting their scalability. When we use Krylov subspace solutions to approximate the cubic-regularized Newton step, our results recover the strongest known convergence guarantees to approximate second-order stationary points of general smooth nonconvex functions.

math.OC↗

arXiv2019

Adversarial Training Can Hurt Generalization

While adversarial training can improve robust accuracy (against an adversary), it sometimes hurts standard accuracy (when there is no adversary). Previous work has studied this tradeoff between standard and robust accuracy, but only in the setting where no predictor performs well on both objectives in the infinite data limit. In this paper, we show that even when the optimal predictor with infinite data performs well on both objectives, a tradeoff can still manifest itself with finite data. Furthermore, since our construction is based on a convex learning problem, we rule out optimization concerns, thus laying bare a fundamental tension between robustness and generalization. Finally, we show that robust self-training mostly eliminates this tradeoff by leveraging unlabeled data.

cs.LG↗

arXiv2019

Lower Bounds for Finding Stationary Points I

We prove lower bounds on the complexity of finding $ε$-stationary points (points $x$ such that $\|\nabla f(x)\| \le ε$) of smooth, high-dimensional, and potentially non-convex functions $f$. We consider oracle-based complexity measures, where an algorithm is given access to the value and all derivatives of $f$ at a query point $x$. We show that for any (potentially randomized) algorithm $\mathsf{A}$, there exists a function $f$ with Lipschitz $p$th order derivatives such that $\mathsf{A}$ requires at least $ε^{-(p+1)/p}$ queries to find an $ε$-stationary point. Our lower bounds are sharp to within constants, and they show that gradient descent, cubic-regularized Newton's method, and generalized $p$th order regularization are worst-case optimal within their natural function classes.

math.OC↗

arXiv2019

A Rank-1 Sketch for Matrix Multiplicative Weights

We show that a simple randomized sketch of the matrix multiplicative weight (MMW) update enjoys (in expectation) the same regret bounds as MMW, up to a small constant factor. Unlike MMW, where every step requires full matrix exponentiation, our steps require only a single product of the form $e^A b$, which the Lanczos method approximates efficiently. Our key technique is to view the sketch as a $\textit{randomized mirror projection}$, and perform mirror descent analysis on the $\textit{expected projection}$. Our sketch solves the online eigenvector problem, improving the best known complexity bounds by $Ω(\log^5 n)$. We also apply this sketch to semidefinite programming in saddle-point form, yielding a simple primal-dual scheme with guarantees matching the best in the literature.

cs.LG↗

arXiv2019

Stochastic (Approximate) Proximal Point Methods: Convergence, Optimality, and Adaptivity

We develop model-based methods for solving stochastic convex optimization problems, introducing the approximate-proximal point, or aProx, family, which includes stochastic subgradient, proximal point, and bundle methods. When the modeling approaches we propose are appropriately accurate, the methods enjoy stronger convergence and robustness guarantees than classical approaches, even though the model-based methods typically add little to no computational overhead over stochastic subgradient methods. For example, we show that improved models converge with probability 1 and enjoy optimal asymptotic normality results under weak assumptions; these methods are also adaptive to a natural class of what we term easy optimization problems, achieving linear convergence under appropriate strong growth conditions on the objective. Our substantial experimental investigation shows the advantages of more accurate modeling over standard subgradient methods across many smooth and non-smooth optimization problems.

math.OC↗

arXiv2019

The importance of better models in stochastic optimization

Standard stochastic optimization methods are brittle, sensitive to stepsize choices and other algorithmic parameters, and they exhibit instability outside of well-behaved families of objectives. To address these challenges, we investigate models for stochastic minimization and learning problems that exhibit better robustness to problem families and algorithmic parameters. With appropriately accurate models---which we call the aProx family---stochastic methods can be made stable, provably convergent and asymptotically optimal; even modeling that the objective is nonnegative is sufficient for this stability. We extend these results beyond convexity to weakly convex objectives, which include compositions of convex losses with smooth functions common in modern machine learning applications. We highlight the importance of robustness and accurate modeling with a careful experimental evaluation of convergence time and algorithm sensitivity.

math.OC↗

arXiv2019

Analysis of Krylov Subspace Solutions of Regularized Nonconvex Quadratic Problems

We provide convergence rates for Krylov subspace solutions to the trust-region and cubic-regularized (nonconvex) quadratic problems. Such solutions may be efficiently computed by the Lanczos method and have long been used in practice. We prove error bounds of the form $1/t^2$ and $e^{-4t/\sqrtκ}$, where $κ$ is a condition number for the problem, and $t$ is the Krylov subspace order (number of Lanczos iterations). We also provide lower bounds showing that our analysis is sharp.

math.OC↗

arXiv2018

Solving (most) of a set of quadratic equalities: Composite optimization for robust phase retrieval

We develop procedures, based on minimization of the composition $f(x) = h(c(x))$ of a convex function $h$ and smooth function $c$, for solving random collections of quadratic equalities, applying our methodology to phase retrieval problems. We show that the prox-linear algorithm we develop can solve phase retrieval problems---even with adversarially faulty measurements---with high probability as soon as the number of measurements $m$ is a constant factor larger than the dimension $n$ of the signal to be recovered. The algorithm requires essentially no tuning---it consists of solving a sequence of convex problems---and it is implementable without any particular assumptions on the measurements taken. We provide substantial experiments investigating our methods, indicating the practical effectiveness of the procedures and showing that they succeed with high probability as soon as $m / n \ge 2$ when the signal is real-valued.

math.ST↗

arXiv2018

Derivative free optimization via repeated classification

We develop an algorithm for minimizing a function using $n$ batched function value measurements at each of $T$ rounds by using classifiers to identify a function's sublevel set. We show that sufficiently accurate classifiers can achieve linear convergence rates, and show that the convergence rate is tied to the difficulty of active learning sublevel sets. Further, we show that the bootstrap is a computationally efficient approximation to the necessary classification scheme. The end result is a computationally efficient derivative-free algorithm requiring no tuning that consistently outperforms other approaches on simulations, standard benchmarks, real-world DNA binding optimization, and airfoil design problems whenever batched function queries are natural.

stat.ML↗

arXiv2017

Unsupervised Transformation Learning via Convex Relaxations

Our goal is to extract meaningful transformations from raw images, such as varying the thickness of lines in handwriting or the lighting in a portrait. We propose an unsupervised approach to learn such transformations by attempting to reconstruct an image from a linear combination of transformations of its nearest neighbors. On handwritten digits and celebrity portraits, we show that even with linear transformations, our method generates visually high-quality modified images. Moreover, since our method is semiparametric and does not model the data distribution, the learned transformations extrapolate off the training data and can be applied to new types of images.

stat.ML↗

arXiv2017

Lower Bounds for Finding Stationary Points II: First-Order Methods

We establish lower bounds on the complexity of finding $ε$-stationary points of smooth, non-convex high-dimensional functions using first-order methods. We prove that deterministic first-order methods, even applied to arbitrarily smooth functions, cannot achieve convergence rates in $ε$ better than $ε^{-8/5}$, which is within $ε^{-1/15}\log\frac{1}ε$ of the best known rate for such methods. Moreover, for functions with Lipschitz first and second derivatives, we prove no deterministic first-order method can achieve convergence rates better than $ε^{-12/7}$, while $ε^{-2}$ is a lower bound for functions with only Lipschitz gradient. For convex functions with Lipschitz gradient, accelerated gradient descent achieves the rate $ε^{-1}\log\frac{1}ε$, showing that finding stationary points is easier given convexity.

math.OC↗

arXiv2017

Mean Estimation from Adaptive One-bit Measurements

We consider the problem of estimating the mean of a normal distribution under the following constraint: the estimator can access only a single bit from each sample from this distribution. We study the squared error risk in this estimation as a function of the number of samples and one-bit measurements $n$. We consider an adaptive estimation setting where the single-bit sent at step $n$ is a function of both the new sample and the previous $n-1$ acquired bits. For this setting, we show that no estimator can attain asymptotic mean squared error smaller than $π/(2n)+O(n^{-2})$ times the variance. In other words, one-bit restriction increases the number of samples required for a prescribed accuracy of estimation by a factor of at least $π/2$ compared to the unrestricted case. In addition, we provide an explicit estimator that attains this asymptotic error, showing that, rather surprisingly, only $π/2$ times more samples are required in order to attain estimation performance equivalent to the unrestricted case.

math.ST↗

arXiv2017

Multiclass Classification, Information, Divergence, and Surrogate Risk

We provide a unifying view of statistical information measures, multi-way Bayesian hypothesis testing, loss functions for multi-class classification problems, and multi-distribution $f$-divergences, elaborating equivalence results between all of these objects, and extending existing results for binary outcome spaces to more general ones. We consider a generalization of $f$-divergences to multiple distributions, and we provide a constructive equivalence between divergences, statistical information (in the sense of DeGroot), and losses for multiclass classification. A major application of our results is in multi-class classification problems in which we must both infer a discriminant function $γ$---for making predictions on a label $Y$ from datum $X$---and a data representation (or, in the setting of a hypothesis testing problem, an experimental design), represented as a quantizer $\mathsf{q}$ from a family of possible quantizers $\mathsf{Q}$. In this setting, we characterize the equivalence between loss functions, meaning that optimizing either of two losses yields an optimal discriminant and quantizer $\mathsf{q}$, complementing and extending earlier results of Nguyen et. al. to the multiclass case. Our results provide a more substantial basis than standard classification calibration results for comparing different losses: we describe the convex losses that are consistent for jointly choosing a data representation and minimizing the (weighted) probability of error in multiclass classification problems.

math.ST↗

arXiv2017

"Convex Until Proven Guilty": Dimension-Free Acceleration of Gradient Descent on Non-Convex Functions

We develop and analyze a variant of Nesterov's accelerated gradient descent (AGD) for minimization of smooth non-convex functions. We prove that one of two cases occurs: either our AGD variant converges quickly, as if the function was convex, or we produce a certificate that the function is "guilty" of being non-convex. This non-convexity certificate allows us to exploit negative curvature and obtain deterministic, dimension-free acceleration of convergence for non-convex functions. For a function $f$ with Lipschitz continuous gradient and Hessian, we compute a point $x$ with $\|\nabla f(x)\| \le ε$ in $O(ε^{-7/4} \log(1/ ε) )$ gradient and function evaluations. Assuming additionally that the third derivative is Lipschitz, we require only $O(ε^{-5/3} \log(1/ ε) )$ evaluations.

math.OC↗