DADA: Differentiable Automatic Data Augmentation

By Yonggang Li, Guosheng Hu, Yongtao Wang, Timothy M. Hospedales, Neil Martin Robertson, Yongxin Yang. 2020

118 citations recorded by the source corpus.

Data augmentation (DA) techniques aim to increase data variability, and thus train deep networks with better generalisation. The pioneering AutoAugment automated the search for optimal DA policies with reinforcement learning. However, AutoAugment is extremely computationally expensive, limiting its wide applicability. Followup work such as PBA and Fast AutoAugment improved efficiency, but optimization speed remains a bottleneck. In this paper, we propose Differentiable Automatic Data Augmentation (DADA) which dramatically reduces the cost. DADA relaxes the discrete DA policy selection to a differentiable optimization problem via Gumbel-Softmax. In addition, we introduce an unbiased gradient estimator, RELAX, leading to an efficient and effective one-pass optimization strategy to learn an efficient and accurate DA policy. We conduct extensive experiments on CIFAR-10, CIFAR-100, SVHN, and ImageNet datasets. Furthermore, we demonstrate the value of Auto DA in pre-training for downstream detection problems. Results show our DADA is at least one order of magnitude faster than the state-of-the-art while achieving very comparable accuracy.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Optimization & Training, Meta & Continual Learning

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. Sergey Levine
2. Timothy M. Hospedales
3. Amos J. Storkey
4. Chelsea Finn
5. Antreas Antoniou
6. Denis Yarats
7. Rob Fergus
8. Pieter Abbeel

In Computer Vision and Pattern Recognition

1. Xinggang Wang
2. Luc Van Gool
3. Dacheng Tao
4. Timothy M. Hospedales
5. Bernt Schiele
6. Tao A. Xiang
7. Trevor Darrell
8. Abhinav Kumar Gupta

In Optimization & Training

1. Timothy M. Hospedales
2. Denis Yarats
3. Jie Qin
4. Amos J. Storkey
5. Qian Zhang
6. Ilya Kostrikov
7. Antreas Antoniou
8. Wenyu Liu

In Meta & Continual Learning

1. Amos J. Storkey
2. Paul Micaelli
3. Timothy M. Hospedales
4. Antreas Antoniou
5. Jyoti Prakash Sahoo
6. Jie Qin
7. Yisheng Song
8. Ilya Kostrikov

Citations in the corpus

• Meta-Learning in Neural Networks: A Survey (2020, methodology): “on in Eq. 5 in order to maximize validation performance [144]. Since augmentation operations are typically non-differentiable, this requires reinforcement learning [144], discrete gradient-estimators [145], or evolutionary [146] methods. An open question is whether powerful GANbased data augmentatio”
• Automatic Data Augmentation for Generalization in Deep Reinforcement Learning (2020, background): “While there is a body of work on the automatic use of data augmentation [16, 38, 47], these approaches were designed for supervised learning and, as we explain here, cannot be applied to RL without further algorithmic changes.”
• A Comprehensive Survey of Few-shot Learning: Evolution, Applications, Challenges, and Opportunities (2022, methodology): “DADA [57] Automatic generation of enhancement policies Gradient Descent CIFAR-10 "”
• ResizeMix: Mixing Data with Preserved Object Information and True Labels (2020, methodology): “OHL-AA [34] tries to optimize the probability distribution of augmentations, while Faster AA [23] and DADA [32] use the differentiable optimization directly to search the combinations and magnitudes of augmentations, which can save lots of searching cost.”