Description of Generative Adversarial Networks

papers_discussed/Generative Adversarial Networks.md

@@ -0,0 +1,37 @@
+### Title
+
+ Generative Adversarial Networks
+
+### tl;dr
+
+ 2 models are being simultaneously trained :  a generative model that captures data and tries to imitate it and a discriminant 
+model that tries to differentiate whether the data came from the training data or the generative model
+
+### Describe the method
+
+ An adversarial net is applied first. A prior on input noise variables and a second multilayer perceptron is defined. *D(x)* represents 
+the probability that the data input *x* came from data rather than generator's distribution to the data. The *D* is trained to maximise
+the probability of assigning correct labels to both data samples and the sample from generator(*G*). The *G* is simultaneously trained to 
+minimise log(1 − D(G(z))).
+
+![alt text](https://github.com/the-ethan-hunt/storage/blob/master/GANS.JPG)
+
+Advantages- 
+
+- Only backpropagation is needed to obtain gradients.
+
+- Since there is no need to generate different entries in sample sequentially, it's sample generation is much faster than fully visible 
+nets like WaveNet.
+
+- Compared to VAEs, it is easier to use discrete latent variables. Additionally, VAEs produce deterministic bias to optimise a lower bound
+on the log-likelihood.
+
+
+### Any further details
+
+[Laplacian Pyramid of Adversarial Networks](https://arxiv.org/pdf/1506.05751v1.pdf) is an interesting paper that has a network architecture 
+generating high quality images mistaken 40% of the times when evaluated by humans.
+
+### My two cents
+
+I liked the idea of concept of *D* and *G* playing a two layer minimax game with the following value function *V(G,D)*