In the past years, deep learning has shown great promise for practical application. One of the downsides of deep networks is the amount of labeled data they require to solve each task. Transfer learning and Deep Domain Adaptation techniques have been introduced to address the lack of enough labeled data when trying to solve a new task. In Transfer Learning the aim is to transfer the knowledge within the network from one domain with enough labeled data to to another domain with much less labeled data while Deep Domain Adaptation, networks extract the invariant information from two different domains and find a common feature representation on both of them. With the addition of such functionality, one can easily use a similar task that has enough labeled data to solve a new task without the need for massive labeled data. In this research, we provide a thorough investigation of different approaches to transfer learning and specially deep domain adaptation methods on datasets which knowledge needs to be transferred from simulated(synthetic) data to the real world test data.

* Supervisor: Univ.-Prof. Dr. Gerhard Widmer, JKU Linz, Austria
* Dates: 01.01.2022 - Ongoing