arXiv subjects

Lexing Ying

Publications and source records attributed to Lexing Ying.

At least 109 records · Page 6Linked to original sources

arXiv2021

Why resampling outperforms reweighting for correcting sampling bias with stochastic gradients

A data set sampled from a certain population is biased if the subgroups of the population are sampled at proportions that are significantly different from their underlying proportions. Training machine learning models on biased data sets requires correction techniques to compensate for the bias. We consider two commonly-used techniques, resampling and reweighting, that rebalance the proportions of the subgroups to maintain the desired objective function. Though statistically equivalent, it has been observed that resampling outperforms reweighting when combined with stochastic gradient algorithms. By analyzing illustrative examples, we explain the reason behind this phenomenon using tools from dynamical stability and stochastic asymptotics. We also present experiments from regression, classification, and off-policy prediction to demonstrate that this is a general phenomenon. We argue that it is imperative to consider the objective function design and the optimization algorithm together while addressing the sampling bias.

cs.LG↗

arXiv2021

Shrinkage Estimation of Functions of Large Noisy Symmetric Matrices

We study the problem of estimating functions of a large symmetric matrix $A_n$ when we only have access to a noisy estimate $\hat{A}_n=A_n+σZ_n/\sqrt{n}.$ We are interested in the case that $Z_n$ is a Wigner ensemble and suggest an algorithm based on nonlinear shrinkage of the eigenvalues of $\hat{A}_n.$ As an intermediate step we explain how recovery of the spectrum of $A_n$ is possible using only the spectrum of $\hat{A}_n$. Our algorithm has important applications, for example, in solving high-dimensional noisy systems of equations or symmetric matrix denoising. Throughout our analysis we rely on tools from random matrix theory.

math.PR↗

arXiv2021

An efficient dynamical low-rank algorithm for the Boltzmann-BGK equation close to the compressible viscous flow regime

It has recently been demonstrated that dynamical low-rank algorithms can provide robust and efficient approximation to a range of kinetic equations. This is true especially if the solution is close to some asymptotic limit where it is known that the solution is low-rank. A particularly interesting case is the fluid dynamic limit that is commonly obtained in the limit of small Knudsen number. However, in this case the Maxwellian which describes the corresponding equilibrium distribution is not necessarily low-rank; because of this, the methods known in the literature are only applicable to the weakly compressible case. In this paper, we propose an efficient dynamical low-rank integrator that can capture the fluid limit -- the Navier-Stokes equations -- of the Boltzmann-BGK model even in the compressible regime. This is accomplished by writing the solution as $f=Mg$, where $M$ is the Maxwellian and the low-rank approximation is only applied to $g$. To efficiently implement this decomposition within a low-rank framework requires, in the isothermal case, that certain coefficients are evaluated using convolutions, for which fast algorithms are known. Using the proposed decomposition also has the advantage that the rank required to obtain accurate results is significantly reduced compared to the previous state of the art. We demonstrate this by performing a number of numerical experiments and also show that our method is able to capture sharp gradients/shock waves.

math.NA↗

arXiv2021

Hierarchical Interpolative Factorization Preconditioner for Parabolic Equations

This note proposes an efficient preconditioner for solving linear and semi-linear parabolic equations. With the Crank-Nicholson time stepping method, the algebraic system of equations at each time step is solved with the conjugate gradient method, preconditioned with hierarchical interpolative factorization. Stiffness matrices arising in the discretization of parabolic equations typically have large condition numbers, and therefore preconditioning becomes essential, especially for large time steps. We propose to use the hierarchical interpolative factorization as the preconditioning for the conjugate gradient iteration. Computed only once, the hierarchical interpolative factorization offers an efficient and accurate approximate inverse of the linear system. As a result, the preconditioned conjugate gradient iteration converges in a small number of iterations. Compared to other classical exact and approximate factorizations such as Cholesky or incomplete Cholesky, the hierarchical interpolative factorization can be computed in linear time and the application of its inverse has linear complexity. Numerical experiments demonstrate the performance of the method and the reduction of conjugate gradient iterations.

math.NA↗

arXiv2021

Approximate inversion of discrete Fourier integral operators

This paper introduces a factorization for the inverse of discrete Fourier integral operators that can be applied in quasi-linear time. The factorization starts by approximating the operator with the butterfly factorization. Next, a hierarchical matrix representation is constructed for the hermitian matrix arising from composing the Fourier integral operator with its adjoint. This representation is inverted efficiently with a new algorithm based on the hierarchical interpolative factorization. By combining these two factorizations, an approximate inverse factorization for the Fourier integral operator is obtained as a product of $O(\log N)$ sparse matrices of size $N\times N$. The resulting approximate inverse factorization can be used as a direct solver or as a preconditioner. Numerical examples on 1D and 2D Fourier integral operators, including a generalized Radon transform, demonstrate the performance of this new approach.

math.NA↗

arXiv2021

A semigroup method for high dimensional committor functions based on neural network

This paper proposes a new method based on neural networks for computing the high-dimensional committor functions that satisfy Fokker-Planck equations. Instead of working with partial differential equations, the new method works with an integral formulation based on the semigroup of the differential operator. The variational form of the new formulation is then solved by parameterizing the committor function as a neural network. There are two major benefits of this new approach. First, stochastic gradient descent type algorithms can be applied in the training of the committor function without the need of computing any mixed second-order derivatives. Moreover, unlike the previous methods that enforce the boundary conditions through penalty terms, the new method takes into account the boundary conditions automatically. Numerical results are provided to demonstrate the performance of the proposed method.

math.NA↗

arXiv2021

A Simple Multiscale Method for Mean Field Games

This paper proposes a multiscale method for solving the numerical solution of mean field games which accelerates the convergence and addresses the problem of determining the initial guess. Starting from an approximate solution at the coarsest level, the method constructs approximations on successively finer grids via alternating sweeping, which not only allows for the use of classical time marching numerical schemes but also enables applications to both local and nonlocal problems. At each level, numerical relaxation is used to stabilize the iterative process. A second-order discretization scheme is derived for higher-order convergence. Numerical examples are provided to demonstrate the efficiency of the proposed method in both local and nonlocal, 1-dimensional and 2-dimensional cases.

math.NA↗

arXiv2021

How to Learn when Data Reacts to Your Model: Performative Gradient Descent

Performative distribution shift captures the setting where the choice of which ML model is deployed changes the data distribution. For example, a bank which uses the number of open credit lines to determine a customer's risk of default on a loan may induce customers to open more credit lines in order to improve their chances of being approved. Because of the interactions between the model and data distribution, finding the optimal model parameters is challenging. Works in this area have focused on finding stable points, which can be far from optimal. Here we introduce performative gradient descent (PerfGD), which is the first algorithm which provably converges to the performatively optimal point. PerfGD explicitly captures how changes in the model affects the data distribution and is simple to use. We support our findings with theory and experiments.

cs.LG↗

arXiv2021

Top-$k$ eXtreme Contextual Bandits with Arm Hierarchy

Motivated by modern applications, such as online advertisement and recommender systems, we study the top-$k$ extreme contextual bandits problem, where the total number of arms can be enormous, and the learner is allowed to select $k$ arms and observe all or some of the rewards for the chosen arms. We first propose an algorithm for the non-extreme realizable setting, utilizing the Inverse Gap Weighting strategy for selecting multiple arms. We show that our algorithm has a regret guarantee of $O(k\sqrt{(A-k+1)T \log (|\mathcal{F}|T)})$, where $A$ is the total number of arms and $\mathcal{F}$ is the class containing the regression function, while only requiring $\tilde{O}(A)$ computation per time step. In the extreme setting, where the total number of arms can be in the millions, we propose a practically-motivated arm hierarchy model that induces a certain structure in mean rewards to ensure statistical and computational efficiency. The hierarchical structure allows for an exponential reduction in the number of relevant arms for each context, thus resulting in a regret guarantee of $O(k\sqrt{(\log A-k+1)T \log (|\mathcal{F}|T)})$. Finally, we implement our algorithm using a hierarchical linear function class and show superior performance with respect to well-known benchmarks on simulated bandit feedback experiments using extreme multi-label classification datasets. On a dataset with three million arms, our reduction scheme has an average inference time of only 7.9 milliseconds, which is a 100x improvement.

stat.ML↗

arXiv2021

Multi-Level Fine-Tuning: Closing Generalization Gaps in Approximation of Solution Maps under a Limited Budget for Training Data

In scientific machine learning, regression networks have been recently applied to approximate solution maps (e.g., potential-ground state map of Schrödinger equation). In this paper, we aim to reduce the generalization error without spending more time in generating training samples. However, to reduce the generalization error, the regression network needs to be fit on a large number of training samples (e.g., a collection of potential-ground state pairs). The training samples can be produced by running numerical solvers, which takes much time in many applications. In this paper, we aim to reduce the generalization error without spending more time in generating training samples. Inspired by few-shot learning techniques, we develop the Multi-Level Fine-Tuning algorithm by introducing levels of training: we first train the regression network on samples generated at the coarsest grid and then successively fine-tune the network on samples generated at finer grids. Within the same amount of time, numerical solvers generate more samples on coarse grids than on fine grids. We demonstrate a significant reduction of generalization error in numerical experiments on challenging problems with oscillations, discontinuities, or rough coefficients. Further analysis can be conducted in the Neural Tangent Kernel regime and we provide practical estimators to the generalization error. The number of training samples at different levels can be optimized for the smallest estimated generalization error under the constraint of budget for training data. The optimized distribution of budget over levels provides practical guidance with theoretical insight as in the celebrated Multi-Level Monte Carlo algorithm.

math.NA↗

arXiv2020

A Note on Optimization Formulations of Markov Decision Processes

This note summarizes the optimization formulations used in the study of Markov decision processes. We consider both the discounted and undiscounted processes under the standard and the entropy-regularized settings. For each setting, we first summarize the primal, dual, and primal-dual problems of the linear programming formulation. We then detail the connections between these problems and other formulations for Markov decision processes such as the Bellman equation and the policy gradient method.

math.OC↗

arXiv2020

Achieving Adversarial Robustness Requires An Active Teacher

A new understanding of adversarial examples and adversarial robustness is proposed by decoupling the data generator and the label generator (which we call the teacher). In our framework, adversarial robustness is a conditional concept---the student model is not absolutely robust, but robust with respect to the teacher. Based on the new understanding, we claim that adversarial examples exist because the student cannot obtain sufficient information of the teacher from the training data. Various ways of achieving robustness is compared. Theoretical and numerical evidence shows that to efficiently attain robustness, a teacher that actively provides its information to the student may be necessary.

cs.LG↗

arXiv2020

On the gradient flow structure of the isotropic Landau equation

We prove that the isotropic Landau equation equipped with the Coulomb potential introduced by Krieger-Strain and Gualdani-Guillen can be identified with the gradient flow of the entropy in the probability space with respect to a Riemannian metric tensor with nonlocal mobility. We give characterizations of the corresponding geodesics equations and present a convergence rate result by estimating its Hessian operator.

math.AP↗

arXiv2020

Efficient Long-Range Convolutions for Point Clouds

The efficient treatment of long-range interactions for point clouds is a challenging problem in many scientific machine learning applications. To extract global information, one usually needs a large window size, a large number of layers, and/or a large number of channels. This can often significantly increase the computational cost. In this work, we present a novel neural network layer that directly incorporates long-range information for a point cloud. This layer, dubbed the long-range convolutional (LRC)-layer, leverages the convolutional theorem coupled with the non-uniform Fourier transform. In a nutshell, the LRC-layer mollifies the point cloud to an adequately sized regular grid, computes its Fourier transform, multiplies the result by a set of trainable Fourier multipliers, computes the inverse Fourier transform, and finally interpolates the result back to the point cloud. The resulting global all-to-all convolution operation can be performed in nearly-linear time asymptotically with respect to the number of input points. The LRC-layer is a particularly powerful tool when combined with local convolution as together they offer efficient and seamless treatment of both short and long range interactions. We showcase this framework by introducing a neural network architecture that combines LRC-layers with short-range convolutional layers to accurately learn the energy and force associated with a $N$-body potential. We also exploit the induced two-level decomposition and propose an efficient strategy to train the combined architecture with a reduced number of samples.

stat.ML↗

arXiv2020

Maximizing robustness of point-set registration by leveraging non-convexity

Point-set registration is a classical image processing problem that looks for the optimal transformation between two sets of points. In this work, we analyze the impact of outliers when finding the optimal rotation between two point clouds. The presence of outliers motivates the use of least unsquared deviation, which is a non-smooth minimization problem over non-convex domain. We compare approaches based on non-convex optimization over special orthogonal group and convex relaxations. We show that if the fraction of outliers is larger than a certain threshold, any naive convex relaxation fails to recover the ground truth rotation regardless of the sample size and dimension. In contrast, minimizing the least unsquared deviation directly over the special orthogonal group exactly recovers the ground truth rotation for any level of corruption as long as the sample size is large enough. These theoretical findings are supported by numerical simulations.

math.OC↗

arXiv2020

Borrowing From the Future: An Attempt to Address Double Sampling

For model-free reinforcement learning, one of the main difficulty of stochastic Bellman residual minimization is the double sampling problem, i.e., while only one single sample for the next state is available in the model-free setting, two independent samples for the next state are required in order to perform unbiased stochastic gradient descent. We propose new algorithms for addressing this problem based on the idea of borrowing extra randomness from the future. When the transition kernel varies slowly with respect to the state, it is shown that the training trajectory of new algorithms is close to the one of unbiased stochastic gradient descent. Numerical results for policy evaluation in both tabular and neural network settings are provided to confirm the theoretical findings.

math.OC↗

arXiv2020

Distributed-memory $\mathcal{H}$-matrix Algebra I: Data Distribution and Matrix-vector Multiplication

We introduce a data distribution scheme for $\mathcal{H}$-matrices and a distributed-memory algorithm for $\mathcal{H}$-matrix-vector multiplication. Our data distribution scheme avoids an expensive $Ω(P^2)$ scheduling procedure used in previous work, where $P$ is the number of processes, while data balancing is well-preserved. Based on the data distribution, our distributed-memory algorithm evenly distributes all computations among $P$ processes and adopts a novel tree-communication algorithm to reduce the latency cost. The overall complexity of our algorithm is $O\Big(\frac{N \log N}{P} + α\log P + β\log^2 P \Big)$ for $\mathcal{H}$-matrices under weak admissibility condition, where $N$ is the matrix size, $α$ denotes the latency, and $β$ denotes the inverse bandwidth. Numerically, our algorithm is applied to address both two- and three-dimensional problems of various sizes among various numbers of processes. On thousands of processes, good parallel efficiency is still observed.

math.NA↗

arXiv2020

Semidefinite relaxation of multi-marginal optimal transport for strictly correlated electrons in second quantization

We consider the strictly correlated electron (SCE) limit of the fermionic quantum many-body problem in the second-quantized formalism. This limit gives rise to a multi-marginal optimal transport (MMOT) problem. Here the marginal state space for our MMOT problem is the binary set $\{0,1\}$, and the number of marginals is the number $L$ of sites in the model. The costs of storing and computing the exact solution of the MMOT problem both scale exponentially with respect to $L$. We propose an efficient convex relaxation which can be solved by semidefinite programming (SDP). In particular, the semidefinite constraint is only of size $2L\times 2L$. Moreover, the SDP-based method yields an approximation of the dual potential needed to the perform self-consistent field iteration in the so-called Kohn-Sham SCE framework, which, once converged, yields a lower bound for the total energy of the system. We demonstrate the effectiveness of our methods on spinless and spinful Hubbard-type models. Numerical results indicate that our relaxation methods yield tight lower bounds for the optimal cost, in the sense that the error due to the semidefinite relaxation is much smaller than the intrinsic modeling error of the Kohn-Sham SCE method. We also describe how our relaxation methods generalize to arbitrary MMOT problems with pairwise cost functions.

math.OC↗