Regularizing Deep Networks With Semantic Data Augmentation.

Area of Science:

• Deep Learning
• Computer Vision
• Machine Learning

Background:

• Conventional data augmentation techniques (e.g., flipping, rotation) offer limited diversity for deep networks.
• Deep networks learn linearized features, where specific directions in feature space represent meaningful semantic transformations.
• Existing methods struggle to generate diverse augmented samples, impacting model generalization.

Purpose of the Study:

• To propose a novel semantic data augmentation algorithm to enhance deep network regularization.
• To leverage linearized features in deep networks for generating diverse training samples.
• To introduce an efficient algorithm that complements traditional data augmentation approaches.

Main Methods:

• Developed Implicit Semantic Data Augmentation (ISDA), a method utilizing semantic transformations in the feature space.
• Employed a sampling-based approach to efficiently identify semantically meaningful directions for augmentation.
• Derived an upper bound for the expected cross-entropy loss, leading to a robust CE loss minimization.
• Extended ISDA for semi-supervised learning by minimizing the KL-divergence between original and augmented features.

Main Results:

• ISDA consistently improves the generalization performance of popular deep models like ResNets and DenseNets.
• Demonstrated effectiveness across diverse datasets including CIFAR-10, CIFAR-100, SVHN, ImageNet, and Cityscapes.
• Showcased ISDA's applicability in both supervised and semi-supervised learning settings.
• The algorithm adds minimal computational cost to standard training procedures.

Conclusions:

• Implicit Semantic Data Augmentation (ISDA) is an effective technique for enhancing deep network generalization.
• ISDA offers a computationally efficient and versatile approach to data augmentation.
• The method successfully addresses the limitations of conventional, low-level data augmentation strategies.