The Deep Bootstrap Framework: Good Online Learners are Good Offline Generalizers

Abstract

We propose a new framework for reasoning about generalization in deep learning. The core idea is to couple the Real World, where optimizers take stochastic gradient steps on the empirical loss, to an Ideal World, where optimizers take steps on the population loss. This leads to an alternate decomposition of test error into: (1) the Ideal World test error plus (2) the gap between the two worlds. If the gap (2) is universally small, this reduces the problem of generalization in offline learning to the problem of optimization in online learning. We then give empirical evidence that this gap between worlds can be small in realistic deep learning settings, in particular supervised image classification. For example, CNNs generalize better than MLPs on image distributions in the Real World, but this is "because" they optimize faster on the population loss in the Ideal World. This suggests our framework is a useful tool for understanding generalization in deep learning, and lays a foundation for future research in the area.

arXiv commentAccepted to ICLR 2021

References in corpus (20)

• An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
• Denoising Diffusion Probabilistic Models
• HuggingFace's Transformers: State-of-the-art Natural Language Processing
• Language Models are Few-Shot Learners
• Improved Regularization of Convolutional Neural Networks with Cutout
• Wide Residual Networks
• Scaling Laws for Neural Language Models
• DARTS: Differentiable Architecture Search
• A PAC-Bayesian Approach to Spectrally-Normalized Margin Bounds for Neural Networks
• Computing Nonvacuous Generalization Bounds for Deep (Stochastic) Neural Networks with Many More Parameters than Training Data
• Towards Understanding the Role of Over-Parametrization in Generalization of Neural Networks
• Characterizing Implicit Bias in Terms of Optimization Geometry
• Fantastic Generalization Measures and Where to Find Them
• Overfitting or perfect fitting? Risk bounds for classification and regression rules that interpolate
• ThreeDWorld: A Platform for Interactive Multi-Modal Physical Simulation
• To understand deep learning we need to understand kernel learning
• Norm-Based Capacity Control in Neural Networks
• Affinity and Diversity: Quantifying Mechanisms of Data Augmentation
• Distributional Generalization: A New Kind of Generalization
• Towards Learning Convolutions from Scratch

Cited by in corpus (6)

• Explaining Neural Scaling Laws
• Revisiting Model Stitching to Compare Neural Representations
• The Evolution of Out-of-Distribution Robustness Throughout Fine-Tuning
• Understanding Deep Learning via Decision Boundary
• Generative Models as a Data Source for Multiview Representation Learning
• The Three Stages of Learning Dynamics in High-Dimensional Kernel Methods