New method for efficient sketching of gradients and Hessians.
problem Memory constraints in training machine learning models.
method A novel framework for scalable gradient and HVP sketching tailored for modern hardware.
result Theoretical guarantees and practical applications in training data attribution and Hessian spectrum analysis.
Large-scale distributed training of neural networks is often limited by network bandwidth, wherein the communication time overwhelms the local computation time. Motivated by the success of sketching methods in sub-linear/streaming algorithms, we introduce Sketched SGD, an algorithm for carrying out distributed SGD by c…
Develops accelerated methods for optimization using low-dimensional projected-gradient information.
problem Optimization with low-dimensional projected-gradient information and Nesterov acceleration.
method Randomized-subspace Nesterov accelerated gradient methods for smooth convex and strongly convex optimization.
result Established accelerated oracle-complexity guarantees and unified basis for comparing sketch families.
FedSKETCH and FedSKETCHGATE improve privacy and efficiency in federated learning.
problem Communication and privacy challenges in federated learning.
method Compression of local gradients using count sketch to protect privacy and reduce communication.
result Sharp convergence guarantees and experimental validation of the methods.
We introduce a new sub-linear space sketch---the Weight-Median Sketch---for learning compressed linear classifiers over data streams while supporting the efficient recovery of large-magnitude weights in the model. This enables memory-limited execution of several statistical analyses over streams, including online featu…
New analysis shows SNG's effectiveness in small samples.
problem Limited understanding of SNG in small data settings.
method Sketch-and-project analysis of SNG.
result Global convergence and rate characterization for SNG.
A new method for estimating large-scale linear models with improved precision.
problem Estimating large-scale linear statistical models efficiently.
method Sequential Least-Squares Estimators with Fast Randomized Sketching (SLSE-FRS), integrating Sketch-and-Solve and Iterative-Sketching methods.
result SLSE-FRS produces high-precision estimators, outperforming state-of-the-art methods.
Collaborative learning allows participants to jointly train a model without data sharing. To update the model parameters, the central server broadcasts model parameters to the clients, and the clients send updating directions such as gradients to the server. While data do not leave a client device, the communicated gra…
New algorithm optimizes positions of CountSketch non-zero entries for better data compression.
problem Optimizing positions of CountSketch non-zero entries for better data compression.
method Learning algorithm that optimizes both values and positions of CountSketch non-zero entries.
result Improves accuracy for low rank approximation and other problems like k-means clustering.
Sketching, a dimensionality reduction technique, has received much attention in the statistics community. In this paper, we study sketching in the context of Newton's method for solving finite-sum optimization problems in which the number of variables and data points are both large. We study two forms of sketching that…
Sketching techniques have become popular for scaling up machine learning algorithms by reducing the sample size or dimensionality of massive data sets, while still maintaining the statistical power of big data. In this paper, we study sketching from an optimization point of view: we first show that the iterative Hessia…
DiPS learns to optimize sketching policies for better recommendation quality.
problem Optimizing sketching policies for long-term user interest prediction in recommender systems.
method Differentiable policy for sketching that learns from training data.
result DiPS requires up to 50% fewer sketch items to achieve the same recommendation quality.
Unified bounds for sketched bilinear forms in machine learning and statistics.
problem Uniform bounds on sketched bilinear forms for modern analyses.
method Generic chaining and new techniques for handling suprema over pairs of sets.
result Improved convergence bounds for sketched Federated Learning and bandit algorithms.
Sketching algorithm finds closest point on convex hull efficiently.
problem Finding the closest point on a convex hull of large datasets.
method Sketching procedure to exploit data structure, gradient project method.
result Faster solution than standard optimization algorithms.
Sketch-GNN reduces GNN training time and memory usage to sublinear scales.
problem Training GNNs on large graphs is computationally expensive and memory-intensive.
method Develops a sketch-based algorithm that trains GNNs on compact sketches of graph adjacency and node embeddings.
result Training time and memory usage grow sublinearly with respect to graph size.
SketchOGD improves memory efficiency for continual learning.
problem Catastrophic forgetting in continual learning.
method Matrix sketching to compress model gradients.
result SketchOGD outperforms existing memory-efficient OGD variants.
SkMM selects data for finetuning by balancing bias and variance.
problem Balancing bias and variance in high-dimensional finetuning.
method Gradient sketching for bias reduction and moment matching for variance reduction.
result Gradient sketching selects samples efficiently and accurately.
ADASAP accelerates GP inference for large datasets.
problem Scaling Gaussian process inference to large datasets.
method Approximate sketch-and-project algorithm for solving linear systems.
result Sketch-and-project rapidly converges to true posterior mean.
We propose and study kernel conjugate gradient methods (KCGM) with random projections for least-squares regression over a separable Hilbert space. Considering two types of random projections generated by randomized sketches and Nyström subsampling, we prove optimal statistical results with respect to variants of norms …
Develops an efficient method for large-scale deep learning problems.
problem Efficiently addressing the computational and memory challenges of natural gradient methods in deep learning.
method Randomized techniques for sketching the empirical Fisher information matrix in neural network layers.
result Global convergence to stationary points and fast linear convergence under the NTK case.
Proposes a method to compare neural networks using feature and gradient vectors.
problem Understanding the behavior of neural networks trained on different datasets.
method Defines a similarity index using feature and gradient vectors, and employs sketching techniques for efficient comparison.
result Demonstrates superior performance in computing similarity of neural networks trained on different datasets.
Sketching and stochastic gradient methods are arguably the most common techniques to derive efficient large scale learning algorithms. In this paper, we investigate their application in the context of nonparametric statistical learning. More precisely, we study the estimator defined by stochastic gradient with mini bat…
Sketch-BERT learns vector sketches using BERT-like self-supervised learning.
problem Lack of effective vector sketch representation for recognition and retrieval tasks.
method Generalized BERT to sketch domain with novel embedding networks and self-supervised sketch gestalt learning.
result Improved performance on sketch recognition, retrieval, and gestalt tasks.
Many popular first-order optimization methods (e.g., Momentum, AdaGrad, Adam) accelerate the convergence rate of deep learning models. However, these algorithms require auxiliary parameters, which cost additional memory proportional to the number of parameters in the model. The problem is becoming more severe as deep l…
Sketchy reduces memory and compute requirements for adaptive regularization in deep learning.
problem Prohibitive memory and running time for adaptive regularization methods in deep learning.
method Low-rank sketching approach using Frequent Directions (FD) to reduce memory and compute requirements.
result Efficient interpolation between resource requirements and degradation in regret guarantees with rank k. A new method improves convergence in low-rank approximation.
problem Efficiently solving large-scale numerical linear algebra problems.
method Error-Powered Sketched Inverse Iteration (EPSI) Method.
result Convergence rate improves at least linearly with sketch size.
We address the statistical and optimization impacts of the classical sketch and Hessian sketch used to approximately solve the Matrix Ridge Regression (MRR) problem. Prior research has quantified the effects of classical sketch on the strictly simpler least squares regression (LSR) problem. We establish that classical …
We present a mechanism to compute a sketch (succinct summary) of how a complex modular deep network processes its inputs. The sketch summarizes essential information about the inputs and outputs of the network and can be used to quickly identify key components and summary statistics of the inputs. Furthermore, the sket…
Localized sketching improves matrix multiplication and ridge regression complexity.
problem Efficiently approximate matrix multiplication and ridge regression with limited data availability.
method Localized sketching matrices for block diagonal structure, reducing sample complexity.
result Localized sketching achieves sample complexity matching global sketching methods.
The paper sharpens the analysis of sketch-and-project methods using randomized singular value decomposition.
problem Improving convergence rates of sketch-and-project methods for solving linear systems and non-linear optimization problems.
method Developing a theoretical framework and new spectral bounds for the expected sketched projection matrix.
result The convergence rate improves linearly with sketch size and even faster with certain spectral decays.
New method reduces linear regret in high-dimensional bandit problems.
problem Heavy spectral tails in streaming matrices lead to linear regret in sketch-based linear bandits.
method Dyadic Block Sketching, a multi-scale matrix sketching approach.
result Achieves sublinear regret bounds without prior knowledge of streaming matrix properties.
This paper concerns a fundamental class of convex matrix optimization problems. It presents the first algorithm that uses optimal storage and provably computes a low-rank approximation of a solution. In particular, when all solutions have low rank, the algorithm converges to a solution. This algorithm, SketchyCGM, modi…
New guarantees for asymmetric sketching in compressive learning.
problem Statistical guarantees for compressive learning with asymmetric feature maps.
method Proves existing guarantees carry over to asymmetric scheme with LPD property, applies to quantized sketches.
result Existing statistical guarantees for compressive learning extend to asymmetric schemes with controlled error.
In this paper, we propose a general framework for sparse and low-rank tensor estimation from cubic sketchings. A two-stage non-convex implementation is developed based on sparse tensor decomposition and thresholded gradient descent, which ensures exact recovery in the noiseless case and stable recovery in the noisy cas…
New analysis proves sketching operators' RIP guarantees for mixture models without importance sampling.
problem Proving sketching operators' Restricted Isometry Property (RIP) for mixture models without assuming importance sampling.
method Proposed alternative analysis based on new deterministic bounds and concentration inequalities.
result Theoretical guarantees for sketching operators without importance sampling.
A fast sketching algorithm solves regularized least squares problems efficiently.
problem Solving large-scale optimization problems with convex or nonconvex regularization.
method Sketching for Regularized Optimization (SRO) algorithm that generates a sketch of the original data matrix and solves the sketched problem.
result General theoretical results for the approximation error between the original and sketched problems, including minimax rates for sparse signal estimation.
Private sketches protect linear regression data privacy.
problem Protecting sensitive information in linear regression.
method Release private sketches of datasets, compute approximate solutions.
result Private sketches maintain good approximation guarantees to the original problem.
Sketching is a randomized dimensionality-reduction method that aims to preserve relevant information in large-scale datasets. Count sketch is a simple popular sketch which uses a randomized hash function to achieve compression. In this paper, we propose a novel extension known as Higher-order Count Sketch (HCS). While …
Paper proposes Nyström sketches for better adaptive compressive learning.
problem Improving adaptability of sketching for compressive learning.
method Data-dependent Nyström approximation for mean embedding.
result Excess risk can be controlled with geometric assumption.
New algorithm reduces sketching dimension to effective problem size.
problem Solving L2-regularized least-squares problems efficiently.
method Randomized algorithm using Gaussian and SRHT embeddings.
result Preserves convergence guarantees with reduced embedding dimension.
Sketching is more fundamental to human cognition than speech. Deep Neural Networks (DNNs) have achieved the state-of-the-art in speech-related tasks but have not made significant development in generating stroke-based sketches a.k.a sketches in vector format. Though there are Variational Auto Encoders (VAEs) for genera…
Density sketches summarize data distributions for accurate sampling and estimation.
problem Accurately estimating and sampling from complex data distributions.
method Online algorithm generating additive density sketches.
result Density sketches provide statistically sound estimators and sampling capabilities.
Improved sketching for logistic and ℓ1 regression with near-linear dimensions.
problem Efficiently approximate ℓ1 and logistic regression problems. method New sketching techniques achieving near-linear dimensions for both problems.
result Achieved near-linear sketching dimensions for ℓ1 and logistic regression. We improve prediction risk estimation for large datasets using sketching and ridge regression.
problem Estimating prediction risks for large datasets efficiently and accurately.
method Random matrix theory, generalized cross validation, sketched ridge regression ensembles, and ensemble trick.
result Consistent risk estimation and prediction intervals for large-scale datasets.
Sketching accelerates structured prediction methods for large datasets.
problem Scaling surrogate kernel methods for large datasets.
method Sketching approximations applied to input and output feature maps.
result Achieves close-to-optimal rates with reduced sketch size.
We introduce a few variants on Frank-Wolfe style algorithms suitable for large scale optimization. We show how to modify the standard Frank-Wolfe algorithm using stochastic gradients, approximate subproblem solutions, and sketched decision variables in order to scale to enormous problems while preserving (up to constan…
Linear regression is a classic method of data analysis. In recent years, sketching -- a method of dimension reduction using random sampling, random projections, or both -- has gained popularity as an effective computational approximation when the number of observations greatly exceeds the number of variables. In this p…
Paper develops a bootstrap method for estimating sketched SVD errors.
problem Lack of tools for accurately estimating sketched SVD errors.
method Develops a fully data-driven bootstrap method for numerical error estimation.
result Allows users to adaptively predict extra work needed for desired error tolerance.