Statistical physics of deep learning
Scientific-Disciplinary Group
02/PHYS-02 - Theoretical Physics Of Fundamental Interactions, Models, Mathematical Methods And Applications, 02/PHYS-06 - Physics For Life Sciences, Environment, And Cultural Heritage, Physics Education And History Of Physics
Description
The project aims to understand the mechanisms of generalization in deep learning using tools from statistical physics. The goal is to develop phenomenological, simplified yet quantitative models that describe the interplay between learning algorithms, data structure, and neural network architectures, strongly informed by empirical observations. Using controlled settings and standard model architectures, we will investigate how and under which conditions learning algorithms select specific solutions among many equivalent ones, thereby inducing implicit biases that are crucial for generalization.The methodology combines quantitative experimentation with mathematical descriptions typical of statistical physics, in order to characterize learning dynamics and the properties of the resulting solutions. The results will contribute to a theoretical understanding of deep learning and may, in the longer term, guide the development of more efficient and robust methods.
Compensation
37,808 Euro
Job posting website
Number of positions
1
Maximum duration
12.0
Funding body
Università di Torino
How to apply
Selection process
Click to expand
View the original posting on the MUR website: Go to MUR website