Deep Learning and Development of Complexity

General artificial intelligence is a grand challenge for humanity and computer scientists in particular. Why this is a grand challenge? There probably were a lot of single celled organisms that turned into a really complex structure that writes, reads such theses, searches for novelty and asks a lot of questions after billions of years of evolution. We used to think that evolution is solved but one can just realise that evolution is an algorithm that created all of nature in the universe with all its diversity and complexity and it becomes not so clear what this algorithm actually is. Most of machine learning algorithms provide a single or a few solutions to the given problem that minimises loss function that reflects how close we are to the known answers and we are really happy with it . The exceptions are generative adversarial networks that learn distributions, generate similar objects and reinforcement learning algorithms that should exist and adapt in environment and maximize its expected reward. The main difference between linear regression and a state of the art generative model is complexity. But it seems like complex things are being learned pretty randomly like humans were not designed for consciousness, philosophy, math, etc. So there is a tough question: can we just directly increase the complexity of some models, generated pictures to avoid tons of meaningless results? In this thesis we will introduce some deep learning concepts, talk about the complexity introduced in \cite{1} and will try to directly increase the complexity of models, generalize them for time-series and add the complexity as an attention-like (\cite{attention}) mechanism for predicting time-series.