Dataset Encoding Platform Using Generative Networks

The amount of data generated each second in the world is colossal, and therefore storage and management of this data in a way which is convenient for common people is becoming increasingly important. With recent advances pushing the limits of computation power to higher levels and Generative networks paving the way for the next disruption in the field of deep learning, we have designed a platform which breaks down an entire dataset into a set of numbers which represent the weights of a neural network which in turn can generate back the whole dataset. We have demonstrated the capabilities of our platform with the NIST dataset by conditional training of a generative network using the Lap1 loss function and achieve a reduction of approximately 1134 times in the amount of data which is required to be stored. We also tested the platform on CIFAR-10 and SVHN datasets and achieved similar results.