A Bayesian Data Augmentation Approach for Learning Deep Models

By Toan Tran, Trung T. Pham, Gustavo Carneiro, Lyle J. Palmer, Ian David Reid. 2017

245 citations recorded by the source corpus.

Data augmentation is an essential part of the training process applied to deep learning models. The motivation is that a robust training process for deep learning models depends on large annotated datasets, which are expensive to be acquired, stored and processed. Therefore a reasonable alternative is to be able to automatically generate new annotated training samples using a process known as data augmentation. The dominant data augmentation approach in the field assumes that new training samples can be obtained via random geometric or appearance transformations applied to annotated training samples, but this is a strong assumption because it is unclear if this is a reliable generative model for producing new training samples. In this paper, we provide a novel Bayesian formulation to data augmentation, where new annotated training points are treated as missing variables and generated based on the distribution learned from the training set. For learning, we introduce a theoretically sound algorithm --- generalised Monte Carlo expectation maximisation, and demonstrate one possible implementation via an extension of the Generative Adversarial Network (GAN). Classification results on MNIST, CIFAR-10 and CIFAR-100 show the better performance of our proposed method compared to the current dominant data augmentation approach mentioned above --- the results also show that our approach produces better classification results than similar GAN models.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Representation & Self-Supervised 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

In Computer Vision and Pattern Recognition

In Representation & Self-Supervised Learning

Citations in the corpus

Randaugment: Practical automated data augmentation with a reduced search space (2019, methodology): “Tran et al. generate augmented data via a Bayesian approach, based on the distribution learned from the training set [48].”
AutoAugment: Learning Augmentation Policies from Data (2018, methodology): “used a Bayesian approach to generate data based on the distribution learned from the training set [42].”
FBNetV3: Joint Architecture-Recipe Search using Neural Acquisition Function (2020, methodology): “These methods use policy search [4], population-based training [12], and Bayesian-based augmentation [35].”
Faster AutoAugment: Learning Augmentation Strategies using Backpropagation (2019, background): “n f1;2;:::;Kg: Augment Xby the kth stage operations: X X#O n=1 [˙ (w k)] nO n) k (X; n) k ;p ( k ) return X Data augmentation can be seen as a process that ﬁlls missing data points in training data [18, 23, 30]. Therefore, we minimize the distance between distributions of the original images and the”
Automatically Learning Data Augmentation Policies for Dialogue Tasks (2019, methodology): “Previous works have also adopted Generative Adversarial Networks (Goodfellow et al., 2014) to either directly generate augmented data (Tran et al., 2017; Sixt et al., 2018; Antoniou et al., 2017; Zhu et al., 2017; Mun et al., 2017; Wang and Perez, 2017), or generate augmentation strategies (Ratner…”
Population Based Augmentation: Efficient Learning of Augmentation Policy Schedules (2019, methodology): “These use methods such as manifold learning (Paschali et al., 2019), Bayesian Optimization (Tran et al., 2017), and generative adversarial networks which generate transformation sequences (Ratner et al., 2017).”
Fast AutoAugment (2019, methodology): “This perspective is also inline with the motivation of Bayesian DA [36].”
Albumentations: fast and flexible image augmentations (2018, background): “Learning more sensible data augmentations for the specific datasets in hand has been explored [14,15,39].”
Learning Data Augmentation Strategies for Object Detection (2019, methodology): “generate augmented data, using a Bayesian approach, based on the distribution learned from the training set [45].”
Effective Data Augmentation With Diffusion Models (2023, unknown): “Training neural networks on synthetic data from generative models was popularized using GANs Antoniou et al. (2017); Tran et al. (2017); Zheng et al. (2017).”
RandAugment: Practical data augmentation with no separate search (2019, methodology): “generate augmented data via a Bayesian approach, based on the distribution learned from the training set [47].”
Tradeoffs in Data Augmentation: An Empirical Study (2021, unknown): “Others have augmented the training set by modelling the data distribution directly (Tran et al., 2017).”
Augment Your Batch: Improving Generalization Through Instance Repetition (2020, methodology): “For example, Bayesian approaches based on the training set distribution [34], generative approaches based on GANs [1, 31] and search methods aim to ﬁnd the best data augmentation policy [4].”
A Comprehensive Survey of Image Augmentation Techniques for Deep Learning (2022, unknown): “Unconditional DCGAN [41], [42, 43, 44] Label-conditional BDA [45], ImbCGAN [46], BAGAN [47], DAGAN [48], MFC-GAN [49], IDA-GAN [50].”
A Survey on Deep Semi-Supervised Learning (2021, background): “D EEP learning has been an active research field with abundant applications in pattern recognition [1], [2], data mining [3], statistical learning [4], computer vision [5], [6], natural language processing [7], [8], etc.”
CLoDSA: a tool for augmentation in classification, localization, detection, semantic segmentation and instance segmentation tasks (2019, background): “This indicates the importance of correctly selecting the set of data augmentation techniques — an active research area [41–43].”
How good is my GAN? (2018, unknown): “Indeed, GANs have been used to train classification networks in a semi-supervised fashion [13, 52] or to facilitate domain adaptation [10].”
Cross-Domain Few-Shot Classification via Learned Feature-Wise Transformation (2020, methodology): “Unlike the hand-crafted approaches such as horizontal ﬂip-ping and random cropping, several recent approaches have been proposed to learn the data augmentation (Cubuk et al., 2019; DeVries & Taylor, 2017a; Lemley et al., 2017; Perez & Wang, 2017; Sixt et al., 2018; Tran et al., 2017).”