Abstract: In typical neural network design, we define an architecture and train it on a dataset. In this talk, we will explore a different approach: constructing neural networks iteratively through a process of composition and addition. By progressively combining simpler networks, we can approximate certain functions very efficiently.
This method relies on the combination of neural network operations — such as addition and composition — and is supported by the Banach Fixed Point Theorem, which provides a mathematical basis for proving convergence of the iterative process. In the talk, we will see simple examples of function approximation using this iterative approach.
Iterative Construction of Neural Networks
09.12.2024 15:30 - 16:30
Organiser:
Vienna School of Mathematics
Location:
TUForMath Room DAEGH18, Freihaus, TU Wien (Wiedner Hauptstraße 8-10)