Dirk Helmrich - Midway doctoral student evaluation in Computer Science

Thesis title:

Parallel & Scalable Machine Learning Approaches for Vision Systems leveraging High Performance Computing Systems with Enhanced Simulation Data.

Doctoral committee:

Morris Riedel - Professor - Head of National Competence for HPC & AI at Julich and University of Iceland.
Ebba Þóra Hvannberg - Professor of Computer Science at the University of Iceland.
Jens Henrik Göbbert - Research Scientist for High Performance Computing and Visualization at Scale bei Forschungszentrum Jülich.
Hanno Scharr - Research Scientist at Forschungszentrum Jülich.

Abstract

Synthetic data is one key point to facilitating the practical use of machine learning algorithms. With biological image processing, especially in realistic field configurations, it is almost impossible to gather the data needed to support the image analysis pipelines needed for research. With virtual scenes and plant visualization, the goal of this project is to address precisely these issues. To be able to do that, the learning algorithms and data generators have to perform at scale, facilitating training analysis, robust output models, and successful integration into practical use. This project focusses on using virtual scenes and modern rendering engines to generate an infinite stream of new data. This does not only open up possibilities on improved model training or even attempting harder problems, but also contains challenges for distributed machine learning techniques as well as the question on how much a learning architecture is able to learn. The model pipeline for this concept is the generation of virtual crop fields for plant health analysis. The output of this PhD thesis will be a distributed synthetic data-based training framework, full photorealistic virtual scenes as well as neural networks for field-scale plant analysis.