Abstract:
In recent years, neural networks (NNs) have been shown to approximate a wide range of functions, making them a highly successful hypothesis class across various fields. In this thesis, we investigate how neural-network-based algorithms can be employed to solve diverse problems. Specifically, we study theoretical guarantees for the solution of inverse problems, classification tasks, and the estimation of frequency-localized functions. In all of these problems, our primary interest lies in approximating solutions from a limited number of data samples. Our results are formulated within the recently introduced framework of Practical Existence Theorems (PETs).
Zoom-Link:
univienna.zoom.us/j/66552729413 Meeting-ID: 665 5272 9413 Kenncode: 831011