Lapo Faggi (University of Florence)
Jan 8, 2020 – 11:00 AM
DIISM, Artificial Intelligence laboratory (room 201), Siena SI
Despite their enormous success, a clear theoretical understanding on why Deep Neural Networks work so well and on how they can efficiently extract specific features is still lacking. Recently, some possible explanations were proposed, based on remarkable analogies between Deep Learning and physics-based conceptual frameworks.
During this seminar, we will try to highlight a possible link between Deep Learning and the Renormalization Group, a fundamental tool in statistical Physics and Quantum Field Theory.
The core idea is looking to the depth of a network as a level of abstraction of the corresponding features. This flow from low level to high level features is reminiscent of the flow from short-distance physics to effective long-distance physics, implemented through the renormalization group.
Thus, we will review some recent proposals aiming at making this possible connection more quantitative and explicit.