Generative adversarial networks (GANs)

Having learned a neural network from data, it can be used for prediction. Since the top layers of the network have been trained in a supervised manner to perform a particular classification or prediction task, the top layers are really useful only for that task. A network trained to detect stop signs is useless for detecting handwritten digits or cats.

A fascinating result is obtained by taking the pre-trained bottom layers and studying what the features they have learned look like. This can be achieved by generating images that activate a certain set of neurons in the bottom layers. Looking at the generated images, we can see what the neural network “thinks” a particular feature looks like, or what an image with a select set of features in it would look like. Some even like to talk about the networks “dreaming” or “hallucinating” images.

To actually generate real looking cats, human faces, or other objects (you’ll get whatever you used as the training data), Ian Goodfellow who currently works at Google Brain, proposed a clever combination of two neural networks. The idea is to let the two networks compete against each other. One of the networks is trained to generate images like the ones in the training data. The other network’s task is to separate images generated by the first network from real images from the training data – it is called the adversarial network, and the whole system is called generative adversarial network or a GAN.

The system trains the two models side by side. In the beginning of the training, the adversarial model has an easy task to tell apart the real images from the training data and the clumsy attempts by the generative model. However, as the generative network slowly gets better and better, the adversarial model has to improve as well, and the cycle continues until eventually the generated images are almost indistinguishable from real ones. The GAN tries to not only reproduce the images in the training data: that would be a way too simple strategy to beat the adversarial network. Rather, the system is trained so that it has to be able to generate new, real-looking images too.