Dottorato in TECNOLOGIE DELL'INFORMAZIONE E DELLE COMUNICAZIONI (ICT)

Titolo della tesi: Embedding-based methods for continual learning in neural networks

Humans and other animals have the extraordinary ability to learn from a new experience and at the same time retain knowledge from past experiences. Not only the learned knowledge is preserved, but it is also used in new scenarios and to learn new skills. One of the goals of Artificial Intelligence is to build agents that incorporate the same principles, that are able to continuously learn from the environment and at the same time create a sophisticated understanding of it, to develop more skills and apply them to new problems, while not forgetting past skills and how to solve past tasks. However, despite this shared goal, little research has been done to address this vision, if we exclude some sporadic work. In fact, current state-of-the-art agents suffer from exposure to new data or operating scenarios which slightly differ from the ones they were trained on. In addition to these problems, the datasets constrain the agents to learn from a fixed set of information and tasks, which cannot lead to the emergence of such autonomous agents. Adaptation capabilities are crucial to building agents that can operate in real-world scenarios, but this aspect of artificial intelligence research has been mostly left out of the most studied fields. In this thesis, we study how these ideas are implemented in machine learning agents, especially deep neural network architectures. We give an exhaustive overview of some problems that these agents can encounter and how these are solved in the literature. Moreover, we propose different approaches to enable the continual learning properties in AI agents, all accompanied by exhaustive experimental evaluations.