[Re] On Adverserial Mixup Resynthesis

Narayanan Elavathur Ranganatha, Aniket Didolkar

Abstract: In this paper we aim to reproduce the experiments of \cite{Beckham2019AdversarialMR}. \cite{Beckham2019AdversarialMR} introduced a new data augmentation technique. They propose a method for interpolating hidden state representations of images to produce an image that belongs to one of the class labels in the dataset. They use adverserial loss for training this system. The feature representations obtained through this approach is tested on classification tasks on various datasets. The datasets used for classification include MNIST, KMNIST and SVHN.

Track: Replicability

NeurIPS Paper Id: https://openreview.net/forum?id=Syx9EIIKdN