The representation theory of neural networks
Marco Armenta
Abstract: In this talk I will present recent applications of representation theory to the study of neural networks in artificial intelligence. First, a neural network can be taken as a representation-like object to which we can apply isomorphisms of quiver representations that preserve what a neural network computes. Second, we can encode the decisions and computations of a neural network on a single sample of data in terms of a stable double-framed thin quiver representation, and since the output of a neural network is independent of the representative in the isomorphism class, it makes sense to consider these "data quiver representations" in a moduli space of stable thin representations.
combinatoricscategory theoryrepresentation theory
Audience: researchers in the topic
( slides )
The TRAC Seminar - Théorie de Représentations et ses Applications et Connections
Organizers: | Thomas Brüstle*, Souheila Hassoun |
*contact for this listing |