Text generation with Generative Adversarial Networks (GANs)[go to overview]
Generative Adversarial Networks (GANs) are part of the family of generative models, which focus on producing samples that follow the probability distribution of the real dataset. This architecture is adversarial because it implies training two neural networks against each other. To make the training process more intuitive, the author of GANs, Ian Goodfellow, compared one network with a counterfeiter trying to create fake money and the other with the police, trying to detect counterfeit money from real one. This architecture gained a lot of interest in the computer vision field, in areas such as: image to image translation, improvement of image resolution and text to image generation. However, it did not have the same impact on the natural language filed due to limitations of the original framework. My presentation investigates these limitations and possible solutions in order to understand and help maximize the advantages of GANs in other areas of research.
31.01.19 - 10:15