Registrations open for Applied Harmonic Analysis and Machine Learning 2022

Registrations are now open for Applied Harmonic Analysis and Machine Learning!

The school consists of three courses on applied harmonic analysis and machine learning, given by leading experts. Graduate students in Mathematics, Physics, Computer Science and Engineering, as well as postdoctoral fellows and young researchers, are welcome.

The course will be held from the 5th to the 9th of September at UniGe | DIMA.

Course title: Geometrical aspects of Deep Learning

Instructor: Joan Bruna

Abstract: In this course we will explore the mathematics of deep learning from a geometric perspective. High-dimensional learning captures several approximation, statistical and computational challenges. We will provide a unified description on the current techniques that attempt to draw the line between positive and negative learning results, and in particular we will explore geometric hypothesis spaces that leverage the physical priors of the learning task.

Course title: A tour of reinforcement learning and applications

Instructors: Agnese Seminara & Alessandro Verri

Abstract: The goal of this course is to provide an introduction to reinforcement learning, striking a balance between mathematical formalism and the discussion of real world applications. We will start from the classical framework of optimal control for dynamical systems and discuss the foundations of dynamic programming algorithms to solve Markov Decision Processes. We will derive the Bellman optimality equation and widely used algorithms to find optimal policies (value and policy iteration, Q learning). We will then add complexity by treating the case where the underlying dynamical system is unknown (either the evolution rule, or the current state of the system or both). Finally, we will discuss recent applications of reinforcement learning to biologically inspired navigation, where real agents interact with a complex environment to reach a desired location.

Course title: From random projections to stochastic gradient descent: the “coin flips” that enable large-scale learning

Instructor: Rachel Ward

Abstract: We will give an overview and history of the theory and applications of several crucial randomized embeddings and algorithms which enable modern machine learning methods to operate at large scale, focusing on the Johnson-Lindenstrauss lemma, the randomized singular value decomposition, and stochastic gradient descent.

Check out all the details at the link below and register by July 31st!