Resolving Class Imbalance using Generative Adversarial Networks

Title	Resolving Class Imbalance using Generative Adversarial Networks
Summary	Resolving Class Imbalance using Generative Adversarial Networks
Keywords	GAN, neural networks, deep learning
TimeFrame	Winter 2018, Spring 2019
References	NIPS 2016 Tutorial on GANs https://arxiv.org/pdf/1701.00160.pdf Effective data generation for imbalanced learning using Conditional Generative Adversarial Networks https://www.researchgate.net/publication/319672232_Effective_data_generation_for_imbalanced_learning_using_Conditional_Generative_Adversarial_Networks BAGAN: Data Augmentation with Balancing GAN https://arxiv.org/abs/1803.09655 InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets https://arxiv.org/pdf/1606.03657.pdf
Prerequisites	Artificial Intelligence and Learning Systems courses; good knowledge of machine learning and neural networks; python programming skills for implementing machine learning algorithms.
Author
Supervisor	Sepideh Pashami, Peter Berck
Level
Status

Huge class imbalance can be a challenge for learning tasks. In such cases, different undersampling and oversampling techniques have been usually used to balance the dataset and compensate for the number of samples in minority class. In this thesis we would like to leverage Generative Adversarial Networks (GANs) for addressing the imbalances in data for classification tasks.

The student expect to perform literature review on GANs specifically different structures such as InfoGAN, Conditional GAN, Auxiliary classifier GANs. They should be able to implement and compare performances of different algorithms using different datasets. The student should be able to compare the results with other undersample and oversampling methods, e.g. SMOTE. Modification of standard approaches for classification task is encouraged. Finally, students should reflect on potential failures and limitations of the methods.