The Low-Rank Simplicity Bias in Deep Networks

Abstract

Modern deep neural networks are highly over-parameterized compared to the data on which they are trained, yet they often generalize remarkably well. A flurry of recent work has asked: why do deep networks not overfit to their training data? In this work, we make a series of empirical observations that investigate and extend the hypothesis that deeper networks are inductively biased to find solutions with lower effective rank embeddings. We conjecture that this bias exists because the volume of functions that maps to low effective rank embedding increases with depth. We show empirically that our claim holds true on finite width linear and non-linear models on practical learning paradigms and show that on natural data, these are often the solutions that generalize well. We then show that the simplicity bias exists at both initialization and after training and is resilient to hyper-parameters and learning methods. We further demonstrate how linear over-parameterization of deep non-linear models can be used to induce low-rank bias, improving generalization performance on CIFAR and ImageNet without changing the modeling capacity.

References in corpus (32)

• Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift
• An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
• PyTorch: An Imperative Style, High-Performance Deep Learning Library
• Neural Architecture Search with Reinforcement Learning
• Gaussian Error Linear Units (GELUs)
• Reconciling modern machine learning practice and the bias-variance trade-off
• Scaling Laws for Neural Language Models
• Neural Tangent Kernel: Convergence and Generalization in Neural Networks
• Exact solutions to the nonlinear dynamics of learning in deep linear neural networks
• Skip-Thought Vectors
• Benign Overfitting in Linear Regression
• Deep Learning Scaling is Predictable, Empirically
• Gradient Descent Provably Optimizes Over-parameterized Neural Networks
• A Closer Look at Memorization in Deep Networks
• Spectral Norm Regularization for Improving the Generalizability of Deep Learning
• Blind Super-Resolution Kernel Estimation using an Internal-GAN
• Qualitatively characterizing neural network optimization problems
• Products of Rectangular Random Matrices: Singular Values and Progressive Scattering
• In Search of the Real Inductive Bias: On the Role of Implicit Regularization in Deep Learning
• Deep learning generalizes because the parameter-function map is biased towards simple functions
• Eigenvalues and Singular Values of Products of Rectangular Gaussian Random Matrices
• The Emergence of Spectral Universality in Deep Networks
• The Pitfalls of Simplicity Bias in Neural Networks
• Implicit Regularization in Deep Learning May Not Be Explainable by Norms
• Neural reparameterization improves structural optimization
• Fast Convergence of Natural Gradient Descent for Overparameterized Neural Networks
• SGD on Neural Networks Learns Functions of Increasing Complexity
• Gradient Starvation: A Learning Proclivity in Neural Networks
• ExpandNets: Linear Over-parameterization to Train Compact Convolutional Networks
• Global Convergence of Adaptive Gradient Methods for An Over-parameterized Neural Network
• Towards Resolving the Implicit Bias of Gradient Descent for Matrix Factorization: Greedy Low-Rank Learning
• Increasing Depth Leads to U-Shaped Test Risk in Over-parameterized Convolutional Networks

Cited by in corpus (7)

• Machine Learning and Deep Learning -- A review for Ecologists
• Increasing Depth Leads to U-Shaped Test Risk in Over-parameterized Convolutional Networks
• Layer Folding: Neural Network Depth Reduction using Activation Linearization
• How Does Adversarial Fine-Tuning Benefit BERT?
• The Difficulty of Passive Learning in Deep Reinforcement Learning
• Implicit Greedy Rank Learning in Autoencoders via Overparameterized Linear Networks
• Generalization by design: Shortcuts to Generalization in Deep Learning