PhD Course: Advanced Deep Learning Models
Deep Learning (DL) is used in different fields of engineering and is proving exceptionally useful for solving problems of extraordinary complexity that until a few years ago could only be addressed by human beings. The translation of texts from one language to another, autonomous driving, image analysis are just some of the problems that can be solved by Deep Learning models with performances comparable or superior to those of human operators. The topic of this course will be the presentation of two advanced Deep Learning models: Graph Neural Networks (GNNs) and Deep Reinforcement Learning (DRL).
Graph Neural Networks are Deep Learning models used when the input data does not have a sequential (e.g. texts) or matrix (e.g. images) structure but can be modeled with a graph. The GNNs allow to associate to each node a data structure that summarizes its properties (embedding) and which is calculated by aggregating and processing the information of the node with that of the neighboring nodes that are at most a certain number of hops away from the node itself. This process can be performed with a message-passing mechanism which is very similar to how Convolutional Neural Networks (CNNs) extract features from images. The embeddings provide an easy way to do node-level, edge-level, and graph-level
prediction tasks. GNNs are very powerful tools and are successfully applied in many different domains such as drug research, fraud detection, route planning and network optimization.
Deep Reinforcement Learning models combine the Reinforcement Learning (RL) paradigm with Deep Learning.
Reinforcement Learning requires an agent to operate in an environment by performing actions that change the state of the environment. The agent receives rewards and penalties and has the goal of maximizing his earnings. The agent's decisions are returned by a Deep Learning model that is trained in order to learn the most convenient actions.
Deep Reinforcement Learning models have achieved super-human performance; as an example, they excel in robotic tasks and can beat human players in competitive games (e.g., Atari, StarCraft, Dota, and Go).
seminar room - 5th floor, 42C cube:
meeting room - 44Z cube:
12 H - 3 CFU