Sahivy R. Gonzalez
GANs is an innovation that has been created recently. The concept is pinning two neural networks against each other. One of them is the Generator which as the name entitles generates new data out of a noise input. The other is the Discriminator which tries to determine if the data is real data or fake generated data from inputs coming from a real data set, and generated data from the Generator.
Essentially, We have created a game of cat and mouse between the two neural networks. Thus, it is important to have in mind that we need both networks to train simultaneously in order for both to learn at the same rate, and neither should overpower the other in order for both to be able to progress their learning.
There are many different types of GANs models nowadays. Some seem to be more promising than others. These models are becoming better, and more complicated by the day. For example, the images you see above are of fake people (meaning these images were generated by a neural network). It is an exciting, yet controversial new technology that we must learn to use correctly.
Discriminative models capture the conditional probability of p(Y|X), where Y is the set labels, and X is the dataset. Unlike Generative models which capture the joint probability of p(X,Y), or just p(X) if there is no labels.
In essence, generative models use the distribution of the data, and will assign a likeliness of a given sample. Meanwhile, discriminative models only gives you the likeliness of a label belonging to a given sample.
