le 16/03/2023 à 14h
à Centrale Nantes, LS2N, salle du conseil.

Daniel Haider est un étudiant-chercheur, Acoustics Research Institute (ARI) de l’Académie autrichienne des sciences, invité dans l’équipe SIMS de mars à juin 2023.

Title:
Neural Networks from the Viewpoint of Frame Theory and the Injectivity of ReLU-layers

Abstract:
In our work, we foster a frame-theoretical approach to study several aspects of artificial neural networks. In the first half of this talk, we will introduce the general methodology of describing neural networks by frames and how common machine learning techniques and architectures can be interpreted in this setting. The second half of the talk is dedicated to the integration of non-linear neural activation functions into frame theory. In particular, we will present our recent results on the injectivity behavior of ReLU-layers on different input domains, corresponding methods for reconstruction, and a stability analysis. This is joint work with Peter Balazs and Martin Ehler.