Optimization the training algorithms of machine learning using GAN networks

Artificial intelligence has started to be a part of our lives in many aspects over the past few decades. Developing new products without features using artificial intelligence is not reasonable in the contemporary world. It would not be possible if we were not using the deep learning techniques in machine learning algorithms. Traditional machine learning needs clever human design code that transforms raw data into input features for machine learning algorithms. But with deep learning, learning features from raw data directly are possible and this eases the requirement for subject-matter expertise. GANs are very recent advancement in the field of deep learning. They were not presented before 2014. Their capacity and quality of generating are far better than the other generative techniques in machine learning. Their philosophy is based on self-criticizing techniques for automatically learning representation of features. GANs can be used for generating photorealistic images, colorization, turning a simple sketch into a photorealistic image, increasing the resolution of an image, replacing photo defects with realistic patterns, predicting the next frames in a video, data augmentation, generating text, audio etc. data and more. GANs' architecture is very original in deep learning. They are made up of two neural networks that compete during training. Their structures are very clever and interesting but that leads us to very difficult training sessions. GANs are known as difficult to train, prone to failure and very difficult to hyper-tune. In this thesis we focused on the optimization of some of the GANs. Their philosophies are the key reason to difficulties. For this we first explain the potential difficulties of GANs' trainings. After we retrain some known GANs and compare the results. We propose some structural designs and some optimization parameters to achieve better performant GANs.