Eventi / BOOST: BOlogna Orthogonal Summer Term

Manos Kapritsos, University of Michigan

Minicourse: Formal Verification of Software Systems

For the past 60 years, we have been relying on testing for building robust, bug-free software. And yet, despite our best efforts, bugs slip past our test cases all the time, resulting in a number of incidents that range from benign to catastrophic. In this lecture series, you will be introduced to an entirely different way to build robust software: by writing code that is formally proven to be free of bugs. We will cover the basics of formal verification that is required for real, complex systems: specification, building state machines and reasoning about them using inductive invariants.

Bobby Kleinberg, Cornell University

Minicourse: Multi-Armed Bandit Algorithms

Learning and decision-making problems often boil down to a balancing act between exploring new possibilities and exploiting the best known one. For nearly seventy-five years, the multi-armed bandit problem has been the predominant theoretical model for investigating these issues. Applications in machine learning and electronic markets pushed multi-armed bandits to the forefront of computing research in the present century. These lectures will introduce some seminal algorithms for solving multi-armed bandit problems in Bayesian and prior-independent settings, discuss contemporary extensions to contextual bandit problems, and present applications to settings with strategic users.

Adam Klivans, The University of Texas at Austin

Minicourse: Algorithms for Learning with Distribution Shift

Mitigating distribution shift remains one of the major challenges of AI and machine learning.  Training distributions often deviate significantly from test distributions, and pre-trained models are commonly deployed without a precise understanding of these differences.  In such cases, a model may have poor performance with potentially dangerous consequences. 
In this minicourse we will describe the first set of efficient algorithms for learning with distribution shift for broad classes of functions and distributions.  A major obstacle to obtaining these algorithms is the computational intractability of testing even simple properties of distributions.  Instead we will analyze algorithms that are allowed to fail gracefully or reject if a significant distribution shift is detected.  Our topics will touch on many areas of theoretical computer science and machine learning. 

Iacopo Masi, La Sapienza

Seminar: From Adversarial Noise to Diffusion Noise, and back

The rapid growth of AI capabilities has raised concerns about the robustness of models to adversarial perturbations. 
A common approach to improve robustness is adversarial training (AT), which trains models on worst-case inputs. While AT is typically applied to discriminative classifiers, in the first part of the seminar I will show how generative modeling—specifically Energy-Based Models (EBMs)—can shed light on some of its unexplained behaviors. 
In the second part of the seminar, we flip the script: instead of using generative models to understand robustness, we study the robustness of generative models, focusing on diffusion models. Unlike classifiers, AT for diffusion models must preserve equivariance to maintain alignment with the data distribution while making them resilient to outliers, corrupted data, and adversarial attacks.