Adversarial Learning of General Transformations for Data Augmentation

Abstract

Data augmentation (DA) is fundamental against overfitting in large convolutional neural networks, especially with a limited training dataset. In images, DA is usually based on heuristic transformations, like geometric or color transformations. Instead of using predefined transformations, our work learns data augmentation directly from the training data by learning to transform images with an encoder-decoder architecture combined with a spatial transformer network. The transformed images still belong to the same class but are new, more complex samples for the classifier. Our experiments show that our approach is better than previous generative data augmentation methods, and comparable to predefined transformation methods when training an image classifier.

References in corpus (11)

• Conditional Generative Adversarial Nets
• Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms
• Improved Regularization of Convolutional Neural Networks with Cutout
• The Effectiveness of Data Augmentation in Image Classification using Deep Learning
• Conditional Image Synthesis With Auxiliary Classifier GANs
• Smart Augmentation - Learning an Optimal Data Augmentation Strategy
• Triple Generative Adversarial Nets
• A Bayesian Data Augmentation Approach for Learning Deep Models
• Data augmentation instead of explicit regularization
• Dataset Augmentation in Feature Space
• Do deep nets really need weight decay and dropout?

Cited by in corpus (4)

• Model Patching: Closing the Subgroup Performance Gap with Data Augmentation
• Optimizing Millions of Hyperparameters by Implicit Differentiation
• Learning by Analogy: Reliable Supervision from Transformations for Unsupervised Optical Flow Estimation
• Data Augmentation Using Adversarial Training for Construction-Equipment Classification