Just interpolate: kernel ``ridgeless regression can generalize

DOI10.1214/19-AOS1849zbMATH Open1453.68155arXiv1808.00387OpenAlexW3104969455MaRDI QIDQ2196223FDOQ2196223

Authors: Tengyuan Liang, Alexander Rakhlin 

Publication date: 28 August 2020

Published in: The Annals of Statistics (Search for Journal in Brave)

Abstract: In the absence of explicit regularization, Kernel "Ridgeless" Regression with nonlinear kernels has the potential to fit the training data perfectly. It has been observed empirically, however, that such interpolated solutions can still generalize well on test data. We isolate a phenomenon of implicit regularization for minimum-norm interpolated solutions which is due to a combination of high dimensionality of the input data, curvature of the kernel function, and favorable geometric properties of the data such as an eigenvalue decay of the empirical covariance and kernel matrices. In addition to deriving a data-dependent upper bound on the out-of-sample error, we present experimental evidence suggesting that the phenomenon occurs in the MNIST dataset.

Full work available at URL: https://arxiv.org/abs/1808.00387

Recommendations

• Generalization error of minimum weighted norm and kernel interpolation
• Surprises in high-dimensional ridgeless least squares interpolation
• Benign overfitting in linear regression
• Benefit of Interpolation in Nearest Neighbor Algorithms
• Overparameterization and generalization error: weighted trigonometric interpolation

zbMATH Keywords

kernel methodsreproducing kernel Hilbert spaceshigh dimensionalityimplicit regularizationdata-dependent boundsminimum-norm interpolationspectral decay

Mathematics Subject Classification ID

Nonparametric regression and quantile regression (62G08) Learning and adaptive systems in artificial intelligence (68T05) Hilbert spaces with reproducing kernels (= (proper) functional Hilbert spaces, including de Branges-Rovnyak and other structured spaces) (46E22)

Cites Work

• Title not available (Why is that?)
• Regularization networks and support vector machines
• Generalized Cross-Validation as a Method for Choosing a Good Ridge Parameter
• An introduction to support vector machines and other kernel-based learning methods.
• Kernels for vector-valued functions: a review
• Title not available (Why is that?)
• On early stopping in gradient descent learning
• Title not available (Why is that?)
• A distribution-free theory of nonparametric regression
• Title not available (Why is that?)
• On the limit of the largest eigenvalue of the large dimensional sample covariance matrix
• Optimal rates for the regularized least-squares algorithm
• 10.1162/153244303321897690
• The origins of kriging
• Best choices for regularization parameters in learning theory: on the bias-variance problem.
• Model selection for regularized least-squares algorithm in learning theory
• The spectrum of kernel random matrices
• Kernel Ridge Regression
• Learning Theory

Cited In (45)

• Overparameterization and Generalization Error: Weighted Trigonometric Interpolation
• On the Inconsistency of Kernel Ridgeless Regression in Fixed Dimensions
• The interpolation phase transition in neural networks: memorization and generalization under lazy training
• Generalization error of random feature and kernel methods: hypercontractivity and kernel matrix concentration
• Communication-efficient distributed estimator for generalized linear models with a diverging number of covariates
• Diversity Sampling is an Implicit Regularization for Kernel Methods
• A random matrix analysis of random Fourier features: beyond the Gaussian kernel, a precise phase transition, and the corresponding double descent*
• A Unifying Tutorial on Approximate Message Passing
• Training Neural Networks as Learning Data-adaptive Kernels: Provable Representation and Approximation Benefits
• Canonical thresholding for nonsparse high-dimensional linear regression
• A Multi-resolution Theory for Approximating Infinite-p-Zero-n: Transitional Inference, Individualized Predictions, and a World Without Bias-Variance Tradeoff
• Benign overfitting and adaptive nonparametric regression
• Multilevel Fine-Tuning: Closing Generalization Gaps in Approximation of Solution Maps under a Limited Budget for Training Data
• A sieve stochastic gradient descent estimator for online nonparametric regression in Sobolev ellipsoids
• Title not available (Why is that?)
• Deformed semicircle law and concentration of nonlinear random matrices for ultra-wide neural networks
• Fit without fear: remarkable mathematical phenomena of deep learning through the prism of interpolation
• Theoretical issues in deep networks
• For interpolating kernel machines, minimizing the norm of the ERM solution maximizes stability
• HARFE: hard-ridge random feature expansion
• Generalization Error of Minimum Weighted Norm and Kernel Interpolation
• Deep Neural Networks, Generic Universal Interpolation, and Controlled ODEs
• Mehler’s Formula, Branching Process, and Compositional Kernels of Deep Neural Networks
• Convergence analysis for over-parameterized deep learning
• Title not available (Why is that?)
• Benign Overfitting and Noisy Features
• New equivalences between interpolation and SVMs: kernels and structured features
• Kernel approximation: from regression to interpolation
• Title not available (Why is that?)
• Tractability from overparametrization: the example of the negative perceptron
• Benign overfitting in linear regression
• Improved complexities for stochastic conditional gradient methods under interpolation-like conditions
• A Universal Trade-off Between the Model Size, Test Loss, and Training Loss of Linear Predictors
• On the proliferation of support vectors in high dimensions*
• SVRG meets AdaGrad: painless variance reduction
• Learning from non-random data in Hilbert spaces: an optimal recovery perspective
• Deep networks for system identification: a survey
• Deep learning: a statistical viewpoint
• Title not available (Why is that?)
• Title not available (Why is that?)
• Title not available (Why is that?)
• Surprises in high-dimensional ridgeless least squares interpolation
• On the robustness of minimum norm interpolators and regularized empirical risk minimizers
• Linearized two-layers neural networks in high dimension
• A precise high-dimensional asymptotic theory for boosting and minimum-\(\ell_1\)-norm interpolated classifiers

Uses Software

• Scikit
• MNIST

This page was built for publication: Just interpolate: kernel ``ridgeless regression can generalize

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q2196223)