Analysis of numerical methods for optimization problems in machine learning

Projekt: Avhandling

Projektinformation

Beskrivning

This project intends to analyze and construct new effective numerical methods for optimization problems that arise in connection with applications of machine learning.
The fundamental idea is to reformulate the optimization problem to instead find a stationary solution of a differential equation. This is a basic concept but it hasn't been used much in this area. It allows us to apply modern time stepping methods, and analyze them with powerfull techniques from the theory of ordinary and partial (stochastic) differential equations.
The focus will especially be on using the functional analytic framework of maximal monotone operators for performing rigorous mathematical error analyses.
StatusPågående
Gällande start-/slutdatum2019/12/012024/12/01