Psyncgan for Data Generation Application of Partially Syncronized GAN to Image and Caption Data Generation

A common problem that is seen in most practical applications of the general machine learning algorithms is the insufficient amount of data that has been already tagged for the application in hand. The abundance of unsyncronized (untagged) data is growing exponentially, yet its useability for practical applications is limited due to the amount of time it would take to process each data entry and classify it for each specific task. PSyncGAN is a novel approach that is able to suppliment existing sets of data that are either incomplete or insufficient for practical big data and machine learning applications. In our model, we are solving two multi-modal generative problems at the same time, Image captioning and Image Generation from natural language description with a limited amount of synchronized textual descriptions and images. The benefit of such a model is that there is no strict requirement of having synchronized corpora of text and image, but can actually make heavy use of the large amounts of singular data freely available on the internet, thus to be able to train on an almost infinite amount of data, including the data generated by the model itself. The results described below are indeed very promising, and could constitute a deeper research direction. Besides being able to train on the limited amount of training data, in this paper we also show that this model can be used as a basis for various other deep generative methods that are learning data distribution correlations (both symmetric and asymmetric) as a training data extension, by providing endless training examples for each proposed training data pair.