Chair for Technologies and Management of Digital Transformation
Univ. Prof. Dr. Ing. Tobias Meisen
Industrial Transfer Learning
„Our mission is to create scalable, data-efficient, and sustainable deep learning models for the industrial context.“
In the field of production, deep learning methods offer great potential for the development of innovative solutions for optimization or automation. A central challenge is to ensure a sufficiently large data basis - whether through experiments on real machinse or through simulations. In addition, the models must always be trained on new observations as soon as the production process changes fundamentally. In our research area Industrial Transfer Learning, we conduct research on how these challenges can be overcome in the future and how source data from different process domains or process variations can be used to create robust, efficient and sustainable Deep Learning models. The central research areas are transfer learning, lontinual learning and meta-learning.
Transfer learning addresses methods for using already pre-trained deep learning models to learn new tasks faster and more data efficiently. While it is an already established paradigm in computer vision, there is still much need for research in the industrial context. In sim2real transfer learning (simulation to reality), for example, we are researching reinforcement learning methods for controlling industrial robots that are first trained in simulation and then transferred to complex real-world environments.
Continual Learning und Meta-Learning
Continual learning and meta-learning address approaches for the continuous training of neural networks across different tasks. The networks should be able to leverage their knowledge from previous tasks without forgetting and to improve the learning success for future tasks ("learning to learn"). In the industrial context, for example, we are researching continual learning methods for the sustainable use of neural networks in predictive quality scenarios (validated in different use cases such as the injection molding processes for plastic components).
Selected Relevant Publications
Effects of Domain Randomization on Simulation-to-Reality Transfer of Reinforcement Learning Policies for Industrial Robots
International Conference on Artificial Intelligence 2020 (to be published)
Transfer of Hierarchical Reinforcement Learning Structures for Robotic Manipulation Tasks
International Conference on Computational Science and Computational Intelligence (accepted, to be published)
Industrial Transfer Learning: Boosting Machine Learning in Production
2019 IEEE 17th International Conference on Industrial Informatics (INDIN) , page 274--279.
Transfer-Learning: Bridging the Gap between Real and Simulation Data for Machine Learning in Injection MoldingProcedia CIRP, 72:185--190