# Adversarial Learning of General Transformations for Data Augmentation

**Authors:** Saypraseuth Mounsaveng, David Vazquez, Ismail Ben Ayed, Marco Pedersoli  
**arXiv:** [1909.09801](https://arxiv.org/abs/1909.09801)  
**Date:** 2019-09-21  
**Summary page:** https://www.summarizepaper.com/en/paper/1909.09801v1/

## Key points

- The paper addresses overfitting in large CNNs with limited training data
- Data augmentation techniques are commonly used to mitigate overfitting
- Predefined transformations may not capture the full complexity of the dataset
- The authors propose a novel approach to learn data augmentation directly from the training data
- Their approach uses an encoder-decoder architecture combined with a spatial transformer network
- Their method outperforms previous generative data augmentation techniques and achieves comparable results to predefined transformation methods
- Learning data augmentation directly from the training data effectively addresses overfitting in CNNs with limited datasets
- The paper presents a promising approach to enhance data augmentation in CNNs by leveraging adversarial learning and general transformations.

## Summary

The paper titled "Adversarial Learning of General Transformations for Data Augmentation" addresses the issue of overfitting in large convolutional neural networks (CNNs) when training data is limited. Data augmentation (DA) techniques are commonly used to mitigate overfitting by applying heuristic transformations, such as geometric or color transformations, to the training images. However, these predefined transformations may not capture the full complexity of the dataset. In this work, the authors propose a novel approach to learn data augmentation directly from the training data. They employ an encoder-decoder architecture combined with a spatial transformer network to transform images. Unlike traditional DA methods that rely on predefined transformations, their approach learns to generate new and more complex samples within the same class. The experiments conducted by the authors demonstrate that their method outperforms previous generative data augmentation techniques and achieves comparable results to predefined transformation methods when training an image classifier. This suggests that learning data augmentation directly from the training data can effectively address overfitting in CNNs with limited datasets. Overall, this paper presents a promising approach to enhance data augmentation in CNNs by leveraging adversarial learning and general transformations. The findings highlight its potential for improving classification performance and mitigating overfitting in various computer vision tasks.