Memorizing without overfitting: Bias, variance, and interpolation in over-parameterized models

Abstract

The bias-variance trade-off is a central concept in supervised learning. In classical statistics, increasing the complexity of a model (e.g., number of parameters) reduces bias but also increases variance. Until recently, it was commonly believed that optimal performance is achieved at intermediate model complexities which strike a balance between bias and variance. Modern Deep Learning methods flout this dogma, achieving state-of-the-art performance using "over-parameterized models" where the number of fit parameters is large enough to perfectly fit the training data. As a result, understanding bias and variance in over-parameterized models has emerged as a fundamental problem in machine learning. Here, we use methods from statistical physics to derive analytic expressions for bias and variance in two minimal models of over-parameterization (linear regression and two-layer neural networks with nonlinear data distributions), allowing us to disentangle properties stemming from the model architecture and random sampling of data. In both models, increasing the number of fit parameters leads to a phase transition where the training error goes to zero and the test error diverges as a result of the variance (while the bias remains finite). Beyond this threshold, the test error of the two-layer neural network decreases due to a monotonic decrease in \emph{both} the bias and variance in contrast with the classical bias-variance trade-off. We also show that in contrast with classical intuition, over-parameterized models can overfit even in the absence of noise and exhibit bias even if the student and teacher models match. We synthesize these results to construct a holistic understanding of generalization error and the bias-variance trade-off in over-parameterized models and relate our results to random matrix theory.

21 pages (double column), 6 figures, 32 pages of supplemental material (single column)

References in corpus (10)

• Identifying and attacking the saddle point problem in high-dimensional non-convex optimization
• Mean-field message-passing equations in the Hopfield model and its generalizations
• The cavity method at zero temperature
• More Data Can Hurt for Linear Regression: Sample-wise Double Descent
• Understanding overfitting peaks in generalization error: Analytical risk curves for $l_2$ and $l_1$ penalized interpolation
• Understanding Double Descent Requires a Fine-Grained Bias-Variance Decomposition
• What causes the test error? Going beyond bias-variance via ANOVA
• Wide flat minima and optimal generalization in classifying high-dimensional Gaussian mixtures
• The Perturbative Resolvent Method: spectral densities of random matrix ensembles via perturbation theory
• The Geometry of Over-parameterized Regression and Adversarial Perturbations

Cited by in corpus (13)

• Learning without neurons in physical systems
• Learning through atypical "phase transitions" in overparameterized neural networks
• The Physical Effects of Learning
• What causes the test error? Going beyond bias-variance via ANOVA
• Supervised Learning Guarantee for Quantum AdaBoost
• On the Double Descent of Random Features Models Trained with SGD
• Weight fluctuations in (deep) linear neural networks and a derivation of the inverse-variance flatness relation
• The Geometry of Over-parameterized Regression and Adversarial Perturbations
• Double descent in quantum kernel methods
• Simplified derivations for high-dimensional convex learning problems
• Bias-variance decomposition of overparameterized regression with random linear features
• Tradeoff of generalization error in unsupervised learning
• Backdoor attacks on DNN and GBDT -- A Case Study from the insurance domain