Training BatchNorm and Only BatchNorm: On the Expressive Power of Random Features in CNNs

Abstract

A wide variety of deep learning techniques from style transfer to multitask learning rely on training affine transformations of features. Most prominent among these is the popular feature normalization technique BatchNorm, which normalizes activations and then subsequently applies a learned affine transform. In this paper, we aim to understand the role and expressive power of affine parameters used to transform features in this way. To isolate the contribution of these parameters from that of the learned features they transform, we investigate the performance achieved when training only these parameters in BatchNorm and freezing all weights at their random initializations. Doing so leads to surprisingly high performance considering the significant limitations that this style of training imposes. For example, sufficiently deep ResNets reach 82% (CIFAR-10) and 32% (ImageNet, top-5) accuracy in this configuration, far higher than when training an equivalent number of randomly chosen parameters elsewhere in the network. BatchNorm achieves this performance in part by naturally learning to disable around a third of the random features. Not only do these results highlight the expressive power of affine parameters in deep learning, but - in a broader sense - they characterize the expressive power of neural networks constructed simply by shifting and rescaling random features.

Published in ICLR 2021

References in corpus (3)

• Very Deep Convolutional Networks for Large-Scale Image Recognition
• Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift
• A Learned Representation For Artistic Style

Cited by in corpus (24)

• Tiny Machine Learning: Progress and Futures
• Improving robustness against common corruptions by covariate shift adaptation
• Enable Deep Learning on Mobile Devices: Methods, Systems, and Applications
• Pretrained Transformers as Universal Computation Engines
• Efficient Self-supervised Vision Transformers for Representation Learning
• What shapes feature representations? Exploring datasets, architectures, and training
• TinyTL: Reduce Activations, Not Trainable Parameters for Efficient On-Device Learning
• FreezeNet: Full Performance by Reduced Storage Costs
• Partial transfusion: on the expressive influence of trainable batch norm parameters for transfer learning
• Efficient and Private Federated Learning with Partially Trainable Networks
• Robust Processing-In-Memory Neural Networks via Noise-Aware Normalization
• Batch Normalization Orthogonalizes Representations in Deep Random Networks
• Supermasks in Superposition
• Batch Normalization Preconditioning for Neural Network Training
• Towards Foundation Models and Few-Shot Parameter-Efficient Fine-Tuning for Volumetric Organ Segmentation
• Hidden-Fold Networks: Random Recurrent Residuals Using Sparse Supermasks
• RSO: A Gradient Free Sampling Based Approach For Training Deep Neural Networks
• Deep Random Projection Outlyingness for Unsupervised Anomaly Detection
• Masked Language Modeling and the Distributional Hypothesis: Order Word Matters Pre-training for Little
• Optimizing the Whole-life Cost in End-to-end CNN Acceleration
• Open-Set Face Identification on Few-Shot Gallery by Fine-Tuning
• BN-NAS: Neural Architecture Search with Batch Normalization
• Deconstructing the Structure of Sparse Neural Networks
• Training BatchNorm Only in Neural Architecture Search and Beyond