Data Augmentation Generative Adversarial Networks (1711.04340v3)

Abstract: Effective training of neural networks requires much data. In the low-data regime, parameters are underdetermined, and learnt networks generalise poorly. Data Augmentation alleviates this by using existing data more effectively. However standard data augmentation produces only limited plausible alternative data. Given there is potential to generate a much broader set of augmentations, we design and train a generative model to do data augmentation. The model, based on image conditional Generative Adversarial Networks, takes data from a source domain and learns to take any data item and generalise it to generate other within-class data items. As this generative process does not depend on the classes themselves, it can be applied to novel unseen classes of data. We show that a Data Augmentation Generative Adversarial Network (DAGAN) augments standard vanilla classifiers well. We also show a DAGAN can enhance few-shot learning systems such as Matching Networks. We demonstrate these approaches on Omniglot, on EMNIST having learnt the DAGAN on Omniglot, and VGG-Face data. In our experiments we can see over 13% increase in accuracy in the low-data regime experiments in Omniglot (from 69% to 82%), EMNIST (73.9% to 76%) and VGG-Face (4.5% to 12%); in Matching Networks for Omniglot we observe an increase of 0.5% (from 96.9% to 97.4%) and an increase of 1.8% in EMNIST (from 59.5% to 61.3%).

• The paper introduces a novel GAN-based data augmentation method that enhances performance in low-data regimes.
• It demonstrates significant accuracy gains on Omniglot, EMNIST, and VGG-Face, with improvements such as 69% to 82% on Omniglot for 5 samples per class.
• The approach extends to both vanilla classifiers and few-shot learning frameworks, offering practical benefits for diverse applications.

Overview of Data Augmentation Generative Adversarial Networks (DAGANs)

Introduction

The paper Data Augmentation Generative Adversarial Networks addresses a pertinent challenge in the field of deep learning — the overfitting of neural networks in low-data domains. Traditionally, deep neural networks have demonstrated superior performance across various tasks provided ample data. However, in practical scenarios where data is limited, traditional methods fall short, leading to poor generalization. This paper introduces an innovative approach to enhance data augmentation using Generative Adversarial Networks (GANs) tailored for scenarios with limited data availability.

Methodology

Data Augmentation via GANs

Data augmentation techniques typically involve applying transformations like translations, rotations, and adding noise, which are known to preserve class labels. While effective to an extent, this approach is constrained by the limited set of transformations at its disposal. The authors propose a generative model, specifically a Data Augmentation Generative Adversarial Network (DAGAN), that goes beyond these limitations by learning a broader set of transformations from a source domain with ample data. This model is then applied to a target domain with scarce data.

The DAGAN uses an image-conditional GAN framework. It learns a mapping from any given data instance to generate plausible within-class variations that enrich the dataset. Notably, the DAGAN is versatile as it can generate augmentations for novel classes not seen during training, which is particularly beneficial for few-shot learning tasks.

Experiments and Results

The DAGAN model was empirically validated on three distinct datasets: Omniglot, EMNIST, and VGG-Face. The results underscore the efficacy of the DAGAN-enhanced data augmentation in improving classification performance:

1. Omniglot Dataset: The accuracy for vanilla classifiers in the low-data regime improved significantly from 69% to 82% with the use of DAGAN-generated augmentations for 5 samples per class.
2. EMNIST Dataset: Similar enhancements were observed, with accuracy improving from 73.9% to 76% for 15 samples per class.
3. VGG-Face Dataset: The improvement was substantial, from 4.5% to 12% for 5 samples per class, illustrating the utility of DAGANs even for complex tasks involving facial recognition.

Few-Shot Learning Enhancement

The DAGAN was also integrated into few-shot learning pipelines, specifically Matching Networks, to evaluate its potential in extreme low-data scenarios:

1. Omniglot: The few-shot classification accuracy saw an incremental improvement from 96.9% to 97.4%.
2. EMNIST: An increase from 59.5% to 61.3% was reported.

Implications and Future Directions

The introduction of DAGAN provides a significant leap in addressing low-data challenges by learning generalized data augmentation transformations. The findings suggest several implications:

• Practical Applications: In fields requiring high reliability despite limited data, such as medical imaging or custom facial recognition, DAGANs can enhance model performance without extensive data gathering.
• Theoretical Contributions: By leveraging GANs for data augmentation, the paper opens new avenues in meta-learning and transfer learning, where data augmentation strategies are learned and transferred across domains.

Conclusion

The DAGAN framework proposed in this paper demonstrates a robust methodology for augmenting data in low-data regimes, thus improving neural network training efficacy. Its applications extend to both vanilla classifiers and state-of-the-art few-shot learning frameworks, highlighting its versatility. Future developments could explore more complex architectures and further optimization of the generative process to enhance performance in even more diverse and challenging datasets.