Dynamic data management for continuous retraining

Abstract

Managing dynamic datasets intended to serve as training data for a Machine Learning (ML) model often emerges as very challenging, especially when data is often altered iteratively and already existing ML models should pertain to the data. For example, this applies when new data versions arise from either a generated or aggregated extension of an existing dataset a model has already been trained on. In this work, it is investigated on how a model-based approach for these training data concerns can be provided as well as how the complete process, including the resulting training and retraining process of the ML model, can therein be integrated. Hence, model-based concepts and the implementation are devised to cope with the complexity of iterative data management as an enabler for the integration of continuous retraining routines. With Deep Learning techniques becoming technically feasible and massively being developed further over the last decade, MLOps, aiming to establish DevOps tailored to ML projects, gained crucial relevance. Unfortunately, data-management concepts for iteratively growing datasets with retraining capabilities embedded in a model-driven ML development methodology are unexplored to the best of our knowledge. To fill in this gap, this contribution provides such agile data management concepts and integrates them and continuous retraining into the model-driven ML Framework MontiAnna [18]. The new functionality is evaluated in the context of a research project where ML is exploited for the optimal design of lattice structures for crash applications.