Schedule

Abstract

Literature reviews are a critical academic tool used to investigate the state of the art, summarize existing knowledge, and identify research gaps. This lecture is designed to cover all the basics of the literature review research method, including automated literature search, selection criteria, snowballing, data extraction, data synthesis, and tool support. By the end of the session, attendees will be equipped with the fundamental skills necessary to independently conduct systematic literature reviews. This lecture serves as a stepping stone for all those who aspire to contribute to the academic community through rigorous and methodical literature research. In a written exam, it wil be required the formulation of a concise research design for conducting a systematic literature review on a topic of choice.

Schedule

Abstract

The goal of the course is to introduce the concept of Physics Informed Deep Neural Networks (PINN), discuss its implementation from scratch in PyTorch and using advanced ad-hoc developed open-source libraries such as nvidia- modulus for addressing real-world problems in various fields (engineering, physics, oil). Examination: Option a: Discussion of a research work on the topic, selected by the student and accepted by the instructor; it has to be presented orally with a presentation and with a Git repo offering the students implementation of the code. Option b: Resolution of a small research problem discussed jointly with the instructor; presented either orally with a brief presentation or a written essay, and a Git repo.

Schedule

Abstract

Despite the great classification performances provided by nowadays machine learning (ML) models, it can be difficult to employ their predictions in real-world decision-making processes. Indeed, recognition performances of the ML model can be affected by spurious correlations between input features, bias in the data, and imperfect convergence during the training process. These issues may be undetectable just by looking at the recognition performance [1]. Specifically, the ML model can work as a black box thus preventing decision- makers from validating and trusting their results. In this context, XAI approaches can provide some insights to be sure that the ML model is providing the right outcome for the right reason [2]. XAI is also important for the end-user i.e., who is going to be affected by the decision taken with the support of a ML algorithm. Indeed, recent regulations on personal data processing (i.e., the General Data Protection Regulation [3]), do include some level of explainability as a requirement to be met to employ AI in real-world decision-making processes. For the end user, it would be essential to know why a specific result has been obtained (i.e. why a bank does not allow for a loan) and how this result can be modified [4]. Finally, an explanation can also support researchers and domain experts while investigating the data, verifying the assumptions made upon their investigation, and obtaining new insights from the ML model. To address these issues, some insights into the reasoning behind ML models’ outcomes can be provided via different types of explanations. This course focuses on showing what these types are and when it is more convenient to apply one of them. Each lecture includes a little workshop, and the final assessment consists of the presentation and interpretation of the results obtained in each workshop.

Schedule

Abstract

Strategic decision-making refers to identifying the best way to achieve goals and objectives, maximizing systems performance or social welfare. The problem of matching demand and goods is an important problem in society, some examples are: (i) students assignment in schools, (ii) labor markets where workers and firms are matched, and (iii) organ donation, where patients are matched to potential donors. The economics of “matching and market design” has analyzed these problems and improved real-life institutions in recent years, receiving ever increasing interest from many research areas to solve the resources allocation problems, i.e., the challenging problem of properly assigning resources to machines. This course introduces the theory of matching and market design, presents theoretical fundamentals of matching games, and discusses how the theory can be applied to different applications and scenarios, investigating the main classes of matching games. The emphasis will be put on recent advances in the topic and open questions will be presented so that interested students can promptly come to the frontier and begin their own research. Syllabus: - Introduction: a preliminary discussion of how matching theory is applied to real-world problems. - Two-sided matching: basic theory - One-to-one, one-to-many, many-to-many matchings - Gale-Shapley algorithm - Matching with constraints and Students-projects allocation algorithm - Matching with peer effects and externalities - Matching Applications

Schedule

Abstract

This course shows how to derive performance bounds for queueing systems using Deterministic Network Calculus (DNC), a mathematical framework for worst-case performance analysis. First, we introduce the motivations and context where the DNC is used. Then, we introduce the basic concepts, namely arrival and service curves, and we show how to use them to compute performance bounds in an end-to-end perspective. We then move to tandems of network nodes, showing how the framework allows one to compose models of individual nodes to derive bounds. Finally, we discuss some relevant examples of systems and schedulers that can be modeled and studied with DNC. In all these, we first discuss intuitively the mechanisms that affect the worst-case analysis, how these are reflected in the mathematical model, and finally how they can be explored, also for practical uses, using the software library Nancy. Course Contents in brief: - Context and motivation: not just real-time systems require worst-case guarantees - Deterministic Network Calculus fundamentals: arrival and service curves - Delay, backlog and output bounds - The pay-burst-only-once principle - Modeling examples - Computing DNC results using the Nancy library

Schedule

Abstract

Graphs, which can represent patterns along their relationships, are used in several application domains, including, f.i., bioinformatics, physics, chemistry, and social network analysis. Graph Neural Networks (GNNs) are a large class of machine learning models designed to be directly employed on graph data without any preprocessing. This course aims to present modern GNN models, along with the related theory, some applications and the perspective. First, we will introduce the main concepts behind machine learning for graphs and most common models based on neural networks, f.i., GCNs, GraphSAGE, GAT. Moreover, main theoretical results, which describe the expressive power and generalization capability of GNNs, will be presented. Later, we will examine more advanced models and we will examine the basic building modules that are used to design the recent architectures. A particular focus will be dedicated graph generation based on GNNs. Some applications will also be discussed in order to clarify the possible uses of GNNs.