Generative Adversarial Nets

Contribution

This paper introduces a new generative model that overcomes the difficulties faced by older models like Deep Belief Networks, Variational Autoencoders, etc.

Multilayer Perceptron Networks: A multilayer perceptron is a class of feedforward artificial neural network.
Backpropagation: Backpropagation is a method used in artificial neural networks to calculate a gradient that is needed in the calculation of the weights to be used in the network.

This paper proposes the idea of adversarial training, in which the generative network competes with the discriminative network, which leads to both the networks becoming more efficient in their tasks.
The generative network tries to learn a data distribution, and the discriminative network tries to classify between samples from original data and the generative network.

Two multilayer perceptron networks are used as Generator and Discriminator.
The generator takes in noise z and gives as output a sample x=G(z).
Discriminator takes in sample x and gives as output a probability D(x).
D(x) represents whether sample x came from the generator or original data.
Generator tries to maximize the probability of Discriminator making a mistake.
Discriminator tries to maximize the probability of assigning correct labels to the samples.
Only backpropagation is used for obtaining gradients.
After several iterations, the distribution of generated samples will overlap the distribution of original data.
At this point, the discriminator fails to classify, giving output D(x)=1/2.

The networks were trained on datasets like MNIST, the Toronto Face Database, and CIFAR-10.
The authors believe that the results obtained are at least competitive with other models in the literature.