[processed by Isabel Margolis] Data Augmentation for Improved Classification in Neural Networks

Typically, deep neural networks require lots of training data in order to learn their tasks. For example, the well-known ImageNet Image Classification Challenge comes with an average of 1200 training samples for each of the 1000 classes. For smaller-size training sets, the amount of training data is often artificially increased by some form of data augmentation, i.e., input images are scaled, mirrored and cropped differently.

Data augmentation can also be used to make the networks robust to some input variations that appear in natural data. In an approach to make, for example, facial attribute classification more robust to facial misalignment, in AFFACT: Alignment-Free Facial Attribute Classification Technique we have shown that extended data augmentation including rotation, scaling and blurring of training faces enabled the network to waive the otherwise required face alignment step. Additionally, it was shown that data augmentation during testing can improve classification accuracy considerably.

However, such an approach has never been applied to other input images than faces. In this master thesis, different types and combinations of data augmentation techniques for training and evaluating deep networks should be investigated. Approaches for automatically predicting which types of data augmentation are most helpful for a given training example should be developed. Starting with a smaller-scale problem such as CIFAR-10, experiments are extended to use ImageNet and we try to improve the state of the art in image classification.

Requirements

• A reasonable understanding of deep neural networks and how they learn.
• Programming experience in python and a deep learning framework, optimally, pytorch.
• Experience in parallel programming (on the CPU) in python would be great.
• Decent understanding of written English.