Diffusion-based Approaches in Stochastic Optimization

Job vacancy: From now on

Description

Stochastic optimization represents a powerful approach to solving complex
optimization problems by strategically incorporating randomness. When facing
high-dimensional, nonlinear objective problems, the optimization landscape
often contains multiple local optima. To circumvent undesirable local minima,
exploration of diverse regions of the optimization landscape is necessary. Some
algorithms with theoretical guarantees for its certainty of their overall position
require gradient computations, especially in deep learning. Other novel approaches
are based on diffusion to act in the optimization landscape. We aim to adapt the
gradient-based methods with theoretical guarantees to these diffusion-based
approaches.
This position enables insights in state-of-the-art scientific work, which incorporates
involvement in a scientifically demanding task.

The proposed position entails the following tasks:
+ Literature research about stochastic optimization methods for diffusion-based
neural networks
+ Implementing methods / in papers proposed algorithms in Python
+ Visualizing optimization landscapes and simulation results for publications