Evaluation of Oscillatory Activation Functions on MNIST dataset using GANs

Abstract

MNIST is essentially a database of handwritten digits and thus, having a datasetsuch as this with over 70,000 images makes it a great source to perform exper-imentation on. Thus, in this particular paper, we are working on exploring the use and output of the oscillatory activation functions on the MNIST dataset bygenerating a GAN to generate these kind of handwritten digits.The code has been written essentially in Pytorch and we have used a number of oscillatory activation functions and viewed the results. A GAN is a network in which there are 2 neural networks, i.e. the generator and discriminator which are pitted against each other.This is essentially done to enable the generator to generate a fake image and the discriminator to classify that image as real or fake. The training is done accordingly. Thus, viewing the results of oscialltory activation functions in GANs might open us up to new possibilites and help us understand whether the oscillatory activation functions would yield better results or not in the case of GANs.