Unlocking Creativity in Machines: Exploring the Magic of GANs — A Deep Dive into the Paper
GAN (Generative Adversarial Networks) paper outlines the training process for Generative Adversarial Networks (GANs). Here’s the setup:
What is Adversarial Nets Framework?
The G model is pitted against an adversary, the D model. Discriminative model learns to determine if the sample is from the model distribution or the data distribution.
The authors describe the generative model as analogous to a team of counterfeiters, trying to produce fake currency and use it without detection, while the discriminative model is analogous to the police, trying to detect the counterfeit currency.
In trying to compete against each other they get better at doing their bit and improve their methods until the counterfeits are indistinguishable from genuine articles.
They are more focused on estimating probabilities as compared to generating new images
Related Work
This is a pretty dense piece to unpack. It talks about the various approaches that were previously used and their shortcomings.
First it talks about undirected graphical models with latent variables such as RBMs. These models include computing unnormalised potential…