Deep learning needs lots of data for training; however, in some industrial applications, the significant amount of data may not be available, limiting the deep learning approach. Modern techniques like transfer learning and generative adversarial networks show some hope to solve this challenge. The objective of the project is to propose new techniques for deep learning training.
Deep-learning networks are susceptible to butterfly effect wherein small alterations in the input data can point to drastically distinctive outcomes, making the deep learning network inherently volatile. Thus, the output of deep learning network may be controlled by altering its input or by adding noise. Research has shown that it is possible to fool the deep learning network by adding an imperceptible amount of noise in the input.
Generative Adversarial Networks may have potential to solve the text-to-image problem, but there are challenges in using GANs for NLP. Image classification have got benefitted with large mini-batches and one of the open question the question https://distill.pub/2019/gan-open-problems/#batchsize is if they can also help to scale GANs