A Computational Workflow for Interdisciplinary Deep Learning Projects utilizing bwHPC Infrastructure

Abstract

Deep neural networks have the capability to solve complex tasks through accurate function approximation. The process from submitting domain data and defining process requirements to analyzed data consists of multiple steps, disallowing a simplistic straightforward procedure. It follows that one of the core questions is: how does an application development process facilitating interaction between data scientists and domain experts look like? Practically, two connected challenges need to be addressed. Firstly, it requires a solution for handling large amounts of domain-specific data. Secondly, when dealing with complex deep neural networks, it is essential to find a concept of how model training can be designed in an computationally efficient manner. While tailored solutions for addressing these challenges in interdisciplinary deep learning projects exist, a comprehensive and structured approach is missing. Hence, we present a computational workflow to enhance these kinds of projects concerning data handling, integration of cluster computing resources such as bwHPC infrastructure, and development processes. We exemplify our proposal by means of a biomedical image analysis project.