Faster AutoAugment: Learning Augmentation Strategies using Backpropagation

By Ryuichiro Hataya, Jan Zdenek, Kazuki Yoshizoe, Hideki Nakayama. 2019

238 citations recorded by the source corpus.

Data augmentation methods are indispensable heuristics to boost the performance of deep neural networks, especially in image recognition tasks. Recently, several studies have shown that augmentation strategies found by search algorithms outperform hand-made strategies. Such methods employ black-box search algorithms over image transformations with continuous or discrete parameters and require a long time to obtain better strategies. In this paper, we propose a differentiable policy search pipeline for data augmentation, which is much faster than previous methods. We introduce approximate gradients for several transformation operations with discrete parameters as well as the differentiable mechanism for selecting operations. As the objective of training, we minimize the distance between the distributions of augmented data and the original data, which can be differentiated. We show that our method, Faster AutoAugment, achieves significantly faster searching than prior work without a performance drop.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Optimization & Training, Evaluation & Benchmarks

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

1. Christopher R'e
2. Qi Tian
3. Ekin Dogus Cubuk
4. Lingxi Xie
5. Xinggang Wang
6. Ethan Dyer
7. Raphael Gontijo Lopes
8. Sylvia J. Smullin

In Computer Vision and Pattern Recognition

1. Qi Tian
2. Xinggang Wang
3. Lingxi Xie
4. Luc Van Gool
5. Yi Yang
6. Alan L. Yuille
7. Dacheng Tao
8. Wenyu Liu

In Optimization & Training

1. Raphael Gontijo Lopes
2. Christopher R'e
3. Xinggang Wang
4. Ethan Dyer
5. Sylvia J. Smullin
6. Hongyang Zhang
7. Gregory Valiant
8. Sen Wu

In Evaluation & Benchmarks

1. Christopher R'e
2. Raphael Gontijo Lopes
3. Sen Wu
4. Qi Tian
5. Qian Zhang
6. Gregory Valiant
7. Xiaopeng Zhang
8. Wenyu Liu

Citations in the corpus

• A Comprehensive Survey of Image Augmentation Techniques for Deep Learning (2022, methodology): “Reinforcement learning-based AutoAugment [61], Fast AA [62], PBA [63], Faster AA [64], Ran-dAugment [65], MADAO [66], LDA [67], LSSP [68].”
• Affinity and Diversity: Quantifying Mechanisms of Data Augmentation (2020, background): “Lim et al. (2019); Hataya et al. (2019) have found data augmentation policies by minimizing the distance between the distributions of augmented data and clean data.”
• On the Generalization Effects of Linear Transformations in Data Augmentation (2020, background): “Recent work has proposed random sampling (Cubuk et al., 2019) and Bayesian optimization (Hataya et al., 2019) Proceedings of the 37 th International Conference on Machine Learning , Vienna, Austria, PMLR 119, 2020.”
• Circumventing Outliers of AutoAugment with Knowledge Distillation (2020, methodology): “Eqn (1) is a two-stage optimization problem, for which existing approaches either applied reinforcement learning [4,25,17] or weight-sharing methods [13] which are often more efficient.”
• Tradeoffs in Data Augmentation: An Empirical Study (2021, background): “Lim et al. (2019); Hataya et al. (2020) have found augmentation policies by minimizing the distance between the distributions of augmented data and clean data.”
• ResizeMix: Mixing Data with Preserved Object Information and True Labels (2020, methodology): “We compare the results of our method with CutMix [51], and some saliency-guided mixing augmentations [22, 29, 50], as well as some automated augmentation methods [8, 9, 23, 33].”