# Learning Instance-Specific Data Augmentations

**Authors:** Ning Miao, Emile Mathieu, Yann Dubois, Tom Rainforth, Yee Whye Teh, Adam Foster, Hyunjik Kim  
**arXiv:** [2206.00051](https://arxiv.org/abs/2206.00051)  
**Date:** 2022-05-31  
**Summary page:** https://www.summarizepaper.com/en/paper/2206.00051v1/

## Key points

- InstaAug is a method for learning input-specific data augmentations from training data.
- Existing data augmentation methods assume independence between transformations and inputs.
- InstaAug introduces an augmentation module that maps an input to a distribution over transformations.
- The module is trained alongside the base model in a fully end-to-end manner using only the training data.
- Empirical results show that InstaAug learns meaningful augmentations for various transformation classes.
- InstaAug leads to improved performance on supervised and self-supervised tasks compared to other augmentation methods.
- Existing approaches also assume independent generation of transformations and inputs, but restrict the transformation distribution based on domain expertise.
- For general classes of transformations, this assumption can be justified through the noise outsourcing lemma.
- However, for restricted transformation classes such as location-related parameterizations of crops by a CNN, this assumption may not hold.
- Overall, InstaAug provides a novel approach to learning input-specific augmentations that overcome the limitations of assuming independence between transformations and inputs.

## Summary

The paper introduces InstaAug, a method for learning input-specific data augmentations from training data. Existing data augmentation methods assume independence between transformations and inputs, using the same transformation distribution for all instances. InstaAug addresses this issue by introducing an augmentation module that maps an input to a distribution over transformations. This module is trained alongside the base model in a fully end-to-end manner using only the training data. The authors empirically demonstrate that InstaAug learns meaningful augmentations for various transformation classes, leading to improved performance on supervised and self-supervised tasks compared to augmentations that assume input-transformation independence. In terms of related work, existing approaches also assume independent generation of transformations and inputs. They restrict the transformation distribution to specific classes based on domain expertise. For general classes of transformations, this assumption can be justified through the noise outsourcing lemma. However, for restricted transformation classes such as location-related parameterizations of crops by a CNN, this assumption may not hold. Overall, InstaAug provides a novel approach to learning input-specific augmentations that overcome the limitations of assuming independence between transformations and inputs. The method demonstrates improved performance on various tasks and has potential applications in both supervised and self-supervised learning settings.