Lectures

Ozalp Babaoglu, Università di Bologna

Title: The Impact of BSD Unix on Modern Computing and the Internet:  Origins of the Open Software Movement

In this talk, I will discuss the historical significance of Unix, which may be considered the "grandfather of all modern operating systems". Since its creation at Bell Laboratories in the 1970s, Unix has blossomed into a wide family tree with many branches, one of which has come to be known as “BSD Unix”.  Berkeley Software Distribution (BSD) Unix was a fork from the original Bell Laboratories Research Unix and was developed and distributed by the Computer Systems Research Group (CSRG) at the University of California, Berkeley from the late 1970s throughout the 80s.  During my PhD work at UC Berkeley, I was one of the architects of BSD Unix which was a major factor in the rapid growth of the Internet through its built-in TCP/IP stack and has influenced numerous other modern operating systems including FreeBSD, NetBSD, OpenBSD, SunOS, Solaris, Mac OS/X and iOS.  During the 1980s and 90s, the Berkeley version of UNIX became the standard in education and research, garnering development support from the Defense Advanced Research Projects Agency (DARPA) and was notable for introducing virtual memory and inter-networking.  BSD Unix was widely distributed in source form so that others could learn from it and improve it; this style of software distribution has led to the open source movement, of which BSD Unix is now recognized to be one of the earliest examples.

Gianfranco Bilardi, Università di Padova

Title: Chatting about Transformers: an Introduction to Large Language Models

Proposed in 2017, the Transformer has revolutionized the field of Large Language Models (LLMs), leading to applications that have fascinated the public worldwide, as well as raising intriguing scientific and philosophical questions, on the nature of language and knowledge. The basic function of the transformer consists in evaluating the probability distribution of the "next word" in a text, given the sequence of the preceding words. This function can then be extended to language generation, in response to a given "prompt". 

The seminars will present the key ideas of Machine Learning (ML) that have been successfully combined in the transformer, such as tokenization, word and positional embedding, query-key-value attention, feedforward neural networks, back propagation, and gradient descent. The algorithmic and architectural computing requirements will be considered. The wider implications of LLMs will be briefly discussed.

Paolo Boldi, Università di Milano

Title: When Graphs are Large

Graphs are powerful and versatile tools that find countless applications, from social networks to communication systems of various kinds. Handling large graphs poses a number of new challenges: even storing such graphs in main memory cannot be usually attained in naive ways, and calls for more sophisticated approaches.

In this talk, I will touch on some of the techniques that have been proposed to compress and use very large graphs. Besides compression, I will discuss diffusion-based algorithms using probabilistic counters, with two applications: Milgram-like experiments on very large social networks and the computation of distance-based centrality indices.

Valeria Cardellini, Università di Roma Tor Vergata

Title: Elastic Computing: from Cloud to Edge and Beyond

Elasticity is the degree to which a computing system is able to adapt to fluctuating demands by provisioning and de-provisioning resources in an autonomic manner. It represents a distinguishing feature of Cloud systems and services and becomes even more challenging in highly distributed environments such as the Edge. In this talk, I will discuss what elasticity is and how it can be achieved, also considering an architectural point of view. I will also talk about policies to drive elasticity, both from academia and industry.

Nicolò Cesa-Bianchi, Università di Milano and Politecnico di Milano

Title: The Mathematics of Machine Learning

Machine learning is the main driving force behind the current AI revolution. To provide a solid mathematical foundation to learning systems, we must formally characterize what a machine can learn and what is the minimal amount of training data needed to achieve a desired performance. In this talk, we will show some fundamental results concerning the mathematics of machine learning, stressing their potential and limitations.

Michele Colajanni, Università di Bologna

Title: Digital Innovation through Secure and Resilient Services

Innovation through digitalization represents an inevitable and shared strategy. Equally evident is the fragility of a digital world where all processes, services and supply chains are supported by interconnected digital systems. Hence, we should move from a retrospective in which cybersecurity was perceived as an obstacle to business to an ambitious perspective where we should lead organizations towards digital services that embed cybersecurity and resiliency by design. This route will become even more important as more businesses, industries and services will be driven by digital data and AI. It is important to note that similar perspectives on guarantee of trust and service continuity are also received by the most recent European and US norms.

Abe Davis, Cornell University

Title: Computation for Content Creation

Computers have had a tremendous impact on the ways that we create and consume content. Whether that content is text, digital media (e.g., images, video, and audio), or even tangible manufactured objects, digital tools now play major roles in how we build, capture, or develop most of the things we create. This short course will explore many of those roles. The lectures draw heavily from a course by the same name that I teach for computer science graduate students at Cornell. The tentative topics include:

Representing content: How do we represent different types of content? How might the representation that we expose to the human user of a computational tool differ from the internal representation? What should we consider when designing a representation?
What are computers good at?: The value of a computational tool hinges on being able to leverage certain advantages of computation. We will discuss what computers are (and are not) especially good at, and how these strengths are leveraged in computational tools.
Representing natural signals: We will talk a bit about how to represent and manipulate visual and auditory signals. In other words, an introductory preview of some foundational concepts in vision, graphics, and audio.

Barbara Di Camillo, Università di Padova

Title: Machine Learning Applications in Medicine: from Theory to Practice and Back

In this seminar, I will briefly introduce machine learning, explaining the unique characteristics of its applications in medicine and biology. Specifically, I will focus on the ability of algorithms to generalize and identify reproducible biomarkers. I will then explore methodological aspects related to feature selection that facilitate the discovery of robust biomarkers. Additionally, I will emphasize the importance of explainability in ensuring that machine learning models are transparent and their decisions are understandable, which is crucial for their acceptance and trust in the medical and biological fields.

Title: Using Multi-Agent Models to Simulate Tumor Microenvironment

Multi-agent models are simulation systems composed of multiple autonomous entities, called agents, that interact with each other within a common environment. These models are used to represent and analyze complex behaviors and dynamics of systems where multiple actors, or agents, act and react reciprocally. In this seminar, I will demonstrate how they can be used to simulate the tumor microenvironment, where different types of cells have unique properties and behaviors. Each cell is represented by an autonomous agent, showcasing how this approach can reveal emergent properties of the system.  

Matteo Frigo, Google

Title: Anatomy of Cloud File Systems

We discuss the architecture and the techniques employed by real-world large-scale storage systems, with an emphasis on file storage.  We first discuss the general organization of such systems.  We then dive deep into certain problems that need to be solved for the system to work: 1) how to replicate data reliably and consistently; 2) how to maintain transactional invariants across different parts of the system; 3) how to build a scalable ordered map; 4) how to map the file abstraction into the ordered map; and 5) erasure-coding techniques for efficient storage utilization.  In addition, we discuss techniques for managing congestion, which is an underappreciated problem at all layers of the system.

VIttorio Maniezzo, Università di Bologna

Title: Forecasting Rhapsody: Algorithmic Models and Hybrids for Time Series Forecasting

Knowledge of the future has always been a desire of mankind. Ancient approaches were quite diverse, though not very reliable, a diversity that can also be found in current approaches to time series forecasting. The talk will sketch the variety of backgrounds that have led to state-of-the-art forecasting algorithms, ranging from statistics to plain multilayer perceptrons, from nonlinear optimization to transformer-based models. Furthermore, the broad landscape of applications leads to an interest in the design of combined models.

Lorenzo Orecchia, University of Chicago

Title: Simple and Fast: Algorithm Design via Gradient Descent

In this minicourse, I will explore a recent surprising trend in the field of algorithms: the fastest methods for solving problems on discrete structures, such as graphs, are given by simulating continuous dynamics, e.g., dropping a ball in a potential well and diffusing heat in a conductive medium. Topics will include gradient descent and coordinate descent for linear regression problems, basics of spectral graph theory, combinatorial preconditioning, and variational methods for algorithm design.

Alessandro Panconesi, Sapienza Università di Roma

Title: Hilbert, Gödel, Turing: Computers 'R' Us

In a landmark 1936 paper, Alan Turing famously introduced the concept now known as the Turing machine, a mathematical abstraction that rigorously defines the intuitive and yet elusive notion of an algorithm. His paper presented several revolutionary ideas that profoundly and enduringly influenced the development of science and technology, serving as true harbingers of the computer revolution. While the technical definition of Turing machines may be familiar to many computer science students, it is only by considering Turing's ideas within their proper cultural context that we can fully appreciate their elegance and power. In this popular science talk, I will attempt to do just that.

Keshav Pingali, University of Texas at Austin

Title: Machine Learning for the Rest of Us

It is likely that machine learning (ML) will transform the  way we do science and engineering as radically as computers did 50  years ago. Therefore, just as Computer Science (CS) students need to  know programming regardless of their area of specialization, they will  soon need to know ML to stay relevant as the CS field is transformed  by ML. However, most ML presentations are geared for researchers  specializing in ML, and it can be difficult for students in other  areas to extract the key intuitions and ideas in this rapidly evolving  field. In this series of lectures, we use pictures and the very  intuitive notion of paths in directed graphs to explicate the key  ideas in deep neural networks, convolutional neural networks,  recurrent neural networks, and reinforcement learning (RL).

Samantha Riesenfeld, University of Chicago

Title: Of Mice and Men (and Bytes)

Thanks to recent experimental technologies based on DNA sequencing, genomic data have increased dramatically in size, resolution, and biological scope. Together with bigger and novel types of data come new computational and statistical challenges, as well as more ambitious scientific goals. For example, can we use these noisy, high-dimensional data to demystify the inner workings of the mammalian immune system? Pin down the causes of autoimmune diseases? Improve targeted therapies for different cancers? In this minicourse, I will describe the role of data science in answering these questions, including vignettes from my own research. We will cover some of the unique challenges of extracting insights from these data, current computational strategies, and open problems on the horizon.