/GAN

GAN(Generative Adversarial Networks) is a new class of unsupervised machine learning where two convolutional neural networks, the discriminator and GAN compete in a zero sum game framework. Thomas Wang and I worked on this project to explore this new area of ML.

Primary LanguagePython

GAN

Thoams Wang, Penny Brant

For our final project, we want to continue our study of neural networks in CS999 and specifically focus on learning about Generative Adversarial Networks(GAN). GANs are a class of unsupervised machine learning algorithm. It’s implemented by two neural networks, the GAN and the discriminator network contesting with each other in a zero-sum game framework. One of the main uses of GANs is generating photographs that are superficial to human viewers but hold realistic characteristics to image recognizing softwares.

We believe learning about GANs will gain us tremendous insight to neural nets, (specifically those dealing with image recognition), pytorch vs tensorflow, semi-supervised learning algorithms, deep learning, how to implement multiple competing neural nets.

Furthermore, recently Google researchers have developed psychedelic sticker that, when put onto any image, will trick the neural network into thinking it’s a toaster - something apparently false to human viewers. This since then has started a heated debate to the security of image recognition software. For example, how should image recognition software inside a self-driving car deal with possible interference from GANs? We believe learning more about this new field will give us insight into this new problem posed to the world of high tech.