Transfer learning with soft computing models for regression problems.
The basic task is in this strategic research project in collaboration with SCCH is to transfer knowledge from one or a number of original source tasks to one or more similar target tasks with little training data available for each, while achieving robust and consistent models. Thereby, the following related variants for transfer and representation learning should be considered throughout the project:
- Direct (classical) model transfer from one concrete source task to a target task.
- Multi-task learning as a variant of inductive transfer learning.
-
Learning common transferable representations for a set of tasks.
While the former tries to adopt already available models to a concrete new task, the latter assumes that all tasks are available at once and are thus learned simultaneously. Both variants assume that labeled data is available for all tasks and in both variants it is a challenge to transfer model parameters, model structures and/or input feature representations. Model architectures relevant for this work will be specific types of neural networks for deep structured learning, kernel methods, and fuzzy systems with universal approximation and interpretation capabilities.
The developed approaches will be tuned and evaluated on concrete real-world data from industrial projects.