Affinity and Diversity: Quantifying Mechanisms of Data Augmentation

2020/02/20 by Raphael Gontijo-Lopes, Sylvia Smullin, Gontijo-Lopes, Raphael +6 · 16 citations
Computer Science · Mathematics · #Adversarial Robustness in Machine Learning #Domain Adaptation and Few-Shot Learning #Stochastic Gradient Optimization Techniques #cs.CV #cs.LG #stat.ML

paper · pdf · doi:10.48550/arxiv.2002.08973

10 pages, 7 figures

arxiv created 2020/06/04 · arxiv updated 2020/06/08

Abstract

Though data augmentation has become a standard component of deep neural network training, the underlying mechanism behind the effectiveness of these techniques remains poorly understood. In practice, augmentation policies are often chosen using heuristics of either distribution shift or augmentation diversity. Inspired by these, we seek to quantify how data augmentation improves model generalization. To this end, we introduce interpretable and easy-to-compute measures: Affinity and Diversity. We find that augmentation performance is predicted not by either of these alone but by jointly optimizing the two.

Citations

• Auto-Encoding Variational Bayes
• ImageNet classification with deep convolutional neural networks

Cited by

• Data-Efficient Self-Supervised Algorithms for Fine-Grained Birdsong Analysis
• Does Data Augmentation Benefit from Split BatchNorms
• The Deep Bootstrap Framework: Good Online Learners are Good Offline Generalizers
• Revisiting Mixout: An Overlooked Path to Robust Finetuning
• Hard Negative Mixing for Contrastive Learning
• Selecting Data Augmentation for Simulating Interventions
• Viewmaker Networks: Learning Views for Unsupervised Representation Learning
• KeepAugment: A Simple Information-Preserving Data Augmentation Approach
• Local Patch AutoAugment with Multi-Agent Collaboration
• CutDepth:Edge-aware Data Augmentation in Depth Estimation
• Anatomy of Catastrophic Forgetting: Hidden Representations and Task Semantics
• Data Can Speak for Itself: Quality-guided Utilization of Wireless Synthetic Data
• FMix: Enhancing Mixed Sample Data Augmentation
• On Data-centric Myths
• An Empirical Analysis of the Impact of Data Augmentation on Knowledge Distillation
• A Unified Mixture-View Framework for Unsupervised Representation Learning

Related

• Does Data Augmentation Improve Generalization in NLP?
• AugMax: Adversarial Composition of Random Augmentations for Robust Training
• Data Augmentation Revisited: Rethinking the Distribution Gap between Clean and Augmented Data
• Towards Understanding Why Data Augmentation Improves Generalization
• WeMix: How to Better Utilize Data Augmentation